Professional Documents
Culture Documents
Whole
Whole
Whole
nz
ResearchSpace@Auckland
Copyright Statement
The digital copy of this thesis is protected by the Copyright Act 1994 (New
Zealand).
This thesis may be consulted by you, provided you comply with the provisions of
the Act and the following conditions of use:
• Any use you make of these documents or images must be for research or
private study purposes only, and you may not make them available to any
other person.
• Authors control the copyright of their thesis. You will recognise the
author's right to be identified as the author of this thesis, and due
acknowledgement will be made to the author where appropriate.
• You will obtain the author's permission before publishing any material from
their thesis.
In addition to the above conditions, authors give their consent for the digital copy
of their work to be used subject to the conditions specified on the Library Thesis
Consent Form and Deposit Licence.
The Development of a Diagnostic Approach
to Predicting the Probability of Road
Pavement Failure
Megan R. Schlotjes
The research reported in this thesis was jointly supervised by The University of Auckland and
the University of Birmingham under the Universitas 21 joint PhD scheme. The research,
undertaken between March 2009 and November 2012, was primarily carried out at The
University of Auckland. The University of Birmingham hosted the researcher for the period
between May 2010 and August 2011 and for a subsequent research visit in May 2012 to June
2012.
The contents of this thesis are the original work of the researcher, except where appropriately
acknowledged in the text, and no part of this thesis has been submitted for a degree at another
institution.
Megan Schlotjes
i
ii
Abstract
Road maintenance planning, an essential component of road asset management, preserves the
integrity of road networks. Current state of the art pavement management systems exercise
optimisation tools, pavement deterioration models, and intervention criteria to forecast the
future maintenance requirements of a road network. These tools have been utilised to forecast
future maintenance requirements of road networks; however, with this current approach to
pavement management, uncertainties associated with the failure of individual sections of road
may not always be accounted for explicitly, and therefore the susceptibility of a road network
to failure is unknown.
Predicting the probability of the end of life of a road pavement involves wholly
that engineering knowledge can be well represented in data driven models. To this end, this
thesis describes the development of a diagnostic approach that infers engineering knowledge
into computational models, to quantify the probability of failure of road pavements and
To do so, this research developed a number of failure charts that capture engineering
knowledge, such as citing influential failure factors of road pavements including the
knowledge on road pavement failure was obtained from three sources: literature describing
the fundamentals of pavement design and common causes of road failure, expert knowledge
from the industry identifying relationships between failure mechanisms and causes, and a
data analysis to obtain site-specific causes such as road environments and material properties.
Each chart presents a possible failure path, detailing a set of factors contributing to failure.
iii
A comparative study evaluated the performance of five classification modelling approaches
in order to determine the most suitable technique for this research. Based on performance and
user interpretability criteria, the study identified one based on support vector machines as the
most suitable.
The developed prototype system, consisting of a failure system for rutting, fatigue cracking,
and shear, performed well in both the development phase and network testing of the system
utilising data from the New Zealand Long-term Pavement Performance Programme. A case
study focussing on rural New Zealand roads was carried out, which demonstrated the use of
iv
Acknowledgements
I, as the researcher, would like to sincerely thank my supervisors, Dr. Theunis F.P. Henning
(The University of Auckland) and Dr. Michael P.N. Burrow (University of Birmingham), for
whom without the success of this research would not have been possible. I am appreciative of
their time, inspiration and support through both the highs and lows of the PhD journey. I
further acknowledge the input from my co-supervisors, Dr. John St. George (The University
To my family, I thank you for your support applauding encouragement, despite the
misconception over what exactly it is I do. To my Mum in particular, I thank you for your
constant support, reassurance and love over the years. Without you, the journey would not
The valuable input in various stages of the research from Noel Welsh, Fritz Jooste and Derek
Roux is gratefully acknowledged. Their invaluable knowledge and experience got me out of
some very sticky situations. I would also like to thank the asset managers from Southland
District Council, New Zealand, for their time and input in this research.
Finally, without the financial support from the New Zealand Transport Agency and The
University of Auckland, I would not have been fortunate to experience as many opportunities
Maori Proverb
v
vi
Table of Contents
Chapter One
INTRODUCTION.................................................................................................................... 1
1.1 The Context of Predicting Road Failure .................................................................... 1
1.2 Problem Statement ..................................................................................................... 4
1.3 Aim, Objectives, and Scope of the Research ............................................................. 5
1.3.1 Aim of the Research ............................................................................................... 5
1.3.2 Objectives of the Research ..................................................................................... 5
1.3.3 Scope of the Research ............................................................................................ 6
1.4 Structure of the Thesis ............................................................................................... 8
Chapter Two
LITERATURE REVIEW ..................................................................................................... 11
2.1 Introduction .............................................................................................................. 11
2.2 Background to Road Pavement Failure in New Zealand ......................................... 12
2.2.1 Rutting Failure ..................................................................................................... 12
2.2.2 Cracking Failure................................................................................................... 14
2.2.3 Shear Failure ........................................................................................................ 17
2.2.4 Multiple Failures .................................................................................................. 19
2.3 Representation of Failure Knowledge ..................................................................... 19
2.4 Modelling Terminology ........................................................................................... 21
2.4.1 Deterministic versus Probabilistic ....................................................................... 22
2.4.2 Mechanistic versus Empirical .............................................................................. 23
2.4.3 Linear versus Non-Linear .................................................................................... 23
2.4.4 Classification versus Regression.......................................................................... 24
2.5 Performance Modelling of Infrastructure Assets ..................................................... 25
2.5.1 Deterioration Modelling Techniques ................................................................... 25
2.5.2 Predicting End of Life Probabilities .................................................................... 26
2.6 Application and Performance of Classification Methods ........................................ 28
2.6.1 Logistic Regression Studies ................................................................................. 29
2.6.2 Neural Networks Studies ..................................................................................... 31
vii
Table of Contents
Chapter Three
METHODOLOGY ................................................................................................................ 45
3.1 Introduction .............................................................................................................. 45
3.1.1 Conceptual Design ............................................................................................... 45
3.1.2 Research Methodology ........................................................................................ 46
3.2 Research Modules .................................................................................................... 48
3.2.1 Expert Engineering Knowledge and Failure Understanding ............................... 48
3.2.2 Selecting Suitable Classification Techniques ...................................................... 49
3.2.3 Assessing the Performance of Binary Classifiers ................................................ 50
3.2.4 Prototype System Development ........................................................................... 51
3.2.5 System Testing and Applications......................................................................... 52
3.3 Research Datasets .................................................................................................... 53
3.3.1 State Highway LTPP Dataset ............................................................................... 54
3.3.1.1 Traffic Data................................................................................................. 54
3.3.1.2 Pavement Composition Data ...................................................................... 55
3.3.1.3 Pavement Strength Data ............................................................................. 56
3.3.1.4 Environmental Data .................................................................................... 57
3.3.1.5 Subgrade Sensitivity ................................................................................... 57
3.3.1.6 Condition Data............................................................................................ 57
3.3.1.7 Failure Data ................................................................................................ 59
3.3.2 Local Authority LTPP Dataset ............................................................................. 60
3.3.3 Summary of the LTPP Datasets ........................................................................... 60
3.4 Summary of the Methodology ................................................................................. 61
viii
Table of Contents
Chapter Four
UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge .............................................................................. 63
4.1 Introduction .............................................................................................................. 63
4.1.1 Importance of Understanding Road Pavement Failure ........................................ 64
4.1.2 Methodology for Failure Knowledge and Charts ................................................ 65
4.2 Fault Tree Analysis .................................................................................................. 66
4.3 Generic Pavement Failure ........................................................................................ 66
4.3.1 Failure Factors ..................................................................................................... 69
4.4 Development of the Failure Knowledge and Charts ................................................ 70
4.4.1 Review of the Fundamental Failure Factors ........................................................ 70
4.4.2 Expert Knowledge from the Industry .................................................................. 73
4.4.3 Preliminary Data Analysis ................................................................................... 73
4.4.3.1 Data Analysis Methodology ....................................................................... 74
4.4.3.2 Rutting Relationships ................................................................................. 75
4.4.3.3 Cracking Relationships ............................................................................... 79
4.4.3.4 Shear Relationships .................................................................................... 80
4.4.3.5 Summary of Preliminary Data Analysis ..................................................... 81
4.4.3.6 Additional Notable Findings ...................................................................... 81
4.5 Developed Failure Charts ........................................................................................ 82
4.6 Future Use of the Understanding in this Research................................................... 90
4.7 Summary of Understanding Road Pavement Failure............................................... 91
Chapter Five
A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best? ..................................................................................................... 93
5.1 Introduction .............................................................................................................. 93
5.1.1 Methodology for Selecting the Appropriate Technique ...................................... 94
5.1.2 Dataset for the Study ............................................................................................ 97
5.1.2.1 Data Manipulation ...................................................................................... 97
5.2 Selecting the Classification Techniques .................................................................. 99
5.3 Further Scrutiny of Classification Techniques Using Road Pavement Data ......... 102
ix
Table of Contents
Chapter Six
DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure ................................................................................. 139
6.1 Introduction ............................................................................................................ 139
6.1.1 Conceptual Framework ...................................................................................... 140
6.2 Development of the Prototype System .................................................................. 141
6.2.1 Summarising the Components of the System .................................................... 142
6.3 Dataset for the Development of the Prototype System .......................................... 146
6.4 Performance of the System .................................................................................... 147
6.4.1 Assessment of the System .................................................................................. 147
x
Table of Contents
Chapter Seven
NEW ZEALAND CASE STUDY
Practical Applications of the System .................................................................................. 165
7.1 Introduction ............................................................................................................ 165
7.1.1 Methodology for Case Study ............................................................................. 166
7.1.2 Local Authority LTPP Dataset ........................................................................... 167
7.1.3 Performance of the System ................................................................................ 169
7.2 Project Level Applications ..................................................................................... 170
7.2.1 Analysis of the Predictions ................................................................................ 171
7.3 How does the Rural Local Authority LTPP Sections Perform? ............................ 176
7.3.1 Profile of Likelihood of Failure ......................................................................... 176
7.3.2 Pavement Susceptibility to Failure Modes ........................................................ 179
7.3.3 Symptomatic Problems of the Network ............................................................. 180
7.4 Summary of the Case Study................................................................................... 184
Chapter Eight
SUMMARY AND DISCUSSION
A Review of the Research .................................................................................................... 187
8.1 Introduction ............................................................................................................ 187
xi
Table of Contents
Chapter Nine
CONCLUSIONS AND RECOMMENDATIONS............................................................. 211
9.1 Accomplished Work and Main Findings ............................................................... 211
9.2 Recommendations for Implementing the Research ............................................... 214
9.3 Further Work .......................................................................................................... 215
9.3.1 Recommendations to Address the Limitations of the Research ........................ 215
9.3.2 Recommendations for Further Work ................................................................. 218
9.4 Lessons Learnt from this Research ........................................................................ 219
xii
Table of Contents
xiii
xiv
Table of Figures
xv
Table of Figures
xvi
List of Tables
Table 2-1: Summary of Results from Gil and Johnsson (2011) .............................................. 34
Table 2-2: Advantages and Disadvantages of Generative Classifiers ..................................... 38
Table 2-3: Advantages versus Disadvantages of the Classification Techniques ..................... 41
Table 3-1: Variables included in the LTPP Datasets ............................................................... 61
Table 4-1: Fundamental Factors Associated with Pavement Failure....................................... 69
Table 5-1: Evaluation of Modelling Approaches................................................................... 101
Table 5-2: Kernel Methods for Support Vector Machines .................................................... 108
Table 5-3: Factor Combinations of the Factor Groups .......................................................... 120
Table 5-4: Summary of the Accuracy and Misclassification Error Results for Rutting Failure
(Average of 10-Fold Cross-Validation) ..................................................................... 123
Table 5-5: Summary of the Accuracy and Misclassification Error Results for Fatigue
Cracking Failure (Average of 10-Fold Cross-Validation) ......................................... 124
Table 5-6: Summary of the Accuracy and Misclassification Error Results for Shear Failure
(Average of 10-Fold Cross-Validation) ..................................................................... 125
Table 5-7: Results from the Wilcoxon Signed Rank Test for the Accuracy and
Misclassification Error Performance Measures, showing the Acceptance of the Null
Hypothesis (H0).......................................................................................................... 132
Table 5-8: Results from the Wilcoxon Signed Rank Test for the F-Score and Phi Coefficient
Performance Measures, showing the Acceptance of the Null Hypothesis (H0) ........ 133
Table 5-9: Ranking of the Modelling Techniques ................................................................. 135
Table 5-10: Selection of the Most Appropriate Modelling Technique .................................. 136
Table 6-1: Successful Factor Combinations for the Support Vector Machines Technique ... 142
Table 6-2: Factor Combinations of the State Highway LTPP Dataset per Failure
Mechanism ................................................................................................................. 143
Table 6-3: Assessment of the Performance Measures for Rutting Failure
(Average of 10-Fold Cross-Validation) ..................................................................... 149
Table 6-4: Assessment of the Performance Measures for Shear Failure
(Average of 10-Fold Cross-Validation) ..................................................................... 150
Table 6-5: Assessment of the Performance Measures for Fatigue Cracking Failure
(Average of 10-Fold Cross-Validation) ..................................................................... 151
xvii
List of Tables
Table 6-6: Distribution of Failures in the State Highway LTPP Dataset .............................. 157
Table 6-7: Performance Measures for the Conservative and Probability Theory
Approaches ................................................................................................................ 161
Table 7-1: Distribution of Failures in the Local Authority LTPP Dataset ............................ 169
Table 7-2: Accuracy of Predictions, based on the Confusion Matrices ................................. 170
Table 7-3: Inferred Project Level Statistics from the Rural Local Authority LTPP
Dataset........................................................................................................................ 172
Table 7-4: Secondary Failure Predictions of the Rural Local Authority LTPP Dataset ........ 173
Table 8-1: Size of Evaluation Datasets from the Literature .................................................. 198
Table 8-2: Summary of the Factor Combinations included in the Prototype System ........... 205
xviii
Glossary of Terms
Basecourse layer The aggregate layer beneath the wearing surface, which
(base layer) dissipates the induced forces from traffic loadings prior to
Binary The state of an output which can only be either zero (0) or one
(1).
Critical (most The failure path associated with the maximum predicted
probable) failure path probability from the prototype system, as per failure
xix
Glossary of Terms
Dependent variable(s) The output(s) from the model, such that the model is dependent
vehicles.
Excessive traffic Traffic loadings which exceeds that of the design, or an over-
Fatigue cracking The analysis in this research uses the percentage of surface
failure cracked (by any crack type), rate of cracking, and number of
xx
Glossary of Terms
desired output(s).
Local Authority road New Zealand roads which are managed by Local Authorities
These roads are not included in the main trunk road network.
Low volume road Defined in this research as roads that carry a traffic loading of
input(s).
data.
New Zealand.
Rutting failure The analysis in this research uses the rutting depth and rate of
xxi
Glossary of Terms
Shear failure The analysis in this research uses pothole information, shoving,
shear.
State Highway road The main trunk road network of New Zealand that runs the
xxii
Graduate Centre
ClockTower – East Wing
22 Princes Street, Auckland
Co-Authorship Form Phone: +64 9 373 7599 ext 81321
Fax: +64 9 373 7610
Email: postgraduate@auckland.ac.nz
www.postgrad.auckland.ac.nz
This form is to accompany the submission of any PhD that contains research reported in published or
unpublished co-authored work. Please include one copy of this form for each co-authored work.
Completed forms should be included in all copies of your thesis submitted for examination and library
deposit (including digital deposit), following your thesis Abstract.
Please indicate the chapter/section/pages of this thesis that are extracted from a co-authored work and give the title
and publication details or details of submission of the co-authored work.
"The risks associated with pavement failure in New Zealand"
Presented at the NZIHT and NZTA 10th Annual Conference, Rotorua, New Zealand in November 2009. Work from
Chapters 2 and 3 were delivered, as well as the results from the preliminary data analysis (Appendix B).
Nature of contribution
First author and researcher
by PhD candidate
Extent of contribution
80
by PhD candidate (%)
CO-AUTHORS
Name Nature of Contribution
Theunis F.P. Henning Provided a background to the management of pavement assets in New Zealand, and
revised the paper
Certification by Co-Authors
The undersigned hereby certify that:
the above statement correctly reflects the nature and extent of the PhD candidate’s contribution to this
work, and the nature of the contribution of each of the co-authors; and
in cases where the PhD candidate was the lead author of the work that the candidate wrote the text.
This form is to accompany the submission of any PhD that contains research reported in published or
unpublished co-authored work. Please include one copy of this form for each co-authored work.
Completed forms should be included in all copies of your thesis submitted for examination and library
deposit (including digital deposit), following your thesis Abstract.
Please indicate the chapter/section/pages of this thesis that are extracted from a co-authored work and give the title
and publication details or details of submission of the co-authored work.
"Improved network understanding - a diagnostic approach to the risk of pavement failure"
This paper was presented at the 8th International Conference on Managing Pavement Assets (ICMPA) in November
2011. This paper documents work presented in Chapter 4 of this thesis.
Nature of contribution
First author and researcher
by PhD candidate
Extent of contribution
90
by PhD candidate (%)
CO-AUTHORS
Name Nature of Contribution
Theunis F.P. Henning Supervised the research and revised the paper
Michael P.N. Burrow Supervised the research and revised the paper
Certification by Co-Authors
The undersigned hereby certify that:
the above statement correctly reflects the nature and extent of the PhD candidate’s contribution to this
work, and the nature of the contribution of each of the co-authors; and
in cases where the PhD candidate was the lead author of the work that the candidate wrote the text.
This form is to accompany the submission of any PhD that contains research reported in published or
unpublished co-authored work. Please include one copy of this form for each co-authored work.
Completed forms should be included in all copies of your thesis submitted for examination and library
deposit (including digital deposit), following your thesis Abstract.
Please indicate the chapter/section/pages of this thesis that are extracted from a co-authored work and give the title
and publication details or details of submission of the co-authored work.
"Knowing your risk envelope - An enhanced understanding of the health of road networks"
Presented at the 25th Ingenium Annual Conference in June 2012, presenting the work from Chapter 7 in this thesis.
Nature of contribution
First author and researcher
by PhD candidate
Extent of contribution
75
by PhD candidate (%)
CO-AUTHORS
Name Nature of Contribution
Theunis F.P. Henning Provided a background to the road network and assisted in the analysis of the predictions
Certification by Co-Authors
The undersigned hereby certify that:
the above statement correctly reflects the nature and extent of the PhD candidate’s contribution to this
work, and the nature of the contribution of each of the co-authors; and
in cases where the PhD candidate was the lead author of the work that the candidate wrote the text.
This form is to accompany the submission of any PhD that contains research reported in published or
unpublished co-authored work. Please include one copy of this form for each co-authored work.
Completed forms should be included in all copies of your thesis submitted for examination and library
deposit (including digital deposit), following your thesis Abstract.
Please indicate the chapter/section/pages of this thesis that are extracted from a co-authored work and give the title
and publication details or details of submission of the co-authored work.
"The development of a diagnostic approach that predicts failure risk of flexible road pavements"
Presented at the 7th International Conference on Maintenance and Rehabilitation of Pavements and Tecnological
Control (MAIREPAV), Auckland, New Zealand in August 2012. Work from Chapter 6 was published.
Nature of contribution
First author and researcher
by PhD candidate
Extent of contribution
80
by PhD candidate (%)
CO-AUTHORS
Name Nature of Contribution
Michael P.N. Burrow Assisted in the development of the approach and methodology of the paper
Certification by Co-Authors
The undersigned hereby certify that:
the above statement correctly reflects the nature and extent of the PhD candidate’s contribution to this
work, and the nature of the contribution of each of the co-authors; and
in cases where the PhD candidate was the lead author of the work that the candidate wrote the text.
This form is to accompany the submission of any PhD that contains research reported in published or
unpublished co-authored work. Please include one copy of this form for each co-authored work.
Completed forms should be included in all copies of your thesis submitted for examination and library
deposit (including digital deposit), following your thesis Abstract.
Please indicate the chapter/section/pages of this thesis that are extracted from a co-authored work and give the title
and publication details or details of submission of the co-authored work.
"Incorporating engineering knowledge to diagnose structural pavement failure"
This manuscript has been submitted for publication to the Road and Transport Research Journal (ARRB) in June
2012. Work from Chapter Four was presented in this paper.
Nature of contribution
First author and researcher
by PhD candidate
Extent of contribution
90
by PhD candidate (%)
CO-AUTHORS
Name Nature of Contribution
Theunis F.P. Henning Supervised the research and revised the paper
Michael P.N. Burrow Supervised the research and revised the paper
Certification by Co-Authors
The undersigned hereby certify that:
the above statement correctly reflects the nature and extent of the PhD candidate’s contribution to this
work, and the nature of the contribution of each of the co-authors; and
in cases where the PhD candidate was the lead author of the work that the candidate wrote the text.
This form is to accompany the submission of any PhD that contains research reported in published or
unpublished co-authored work. Please include one copy of this form for each co-authored work.
Completed forms should be included in all copies of your thesis submitted for examination and library
deposit (including digital deposit), following your thesis Abstract.
Please indicate the chapter/section/pages of this thesis that are extracted from a co-authored work and give the title
and publication details or details of submission of the co-authored work.
"A methodology to assess the causes and probability of road pavement structural failure"
This paper has been accepted to be presented at the Transportation Research Board (TRB) 92nd Annual Meeting in
January 2013. This paper documents work presented in Chapter 6 of this thesis.
Nature of contribution
First author and researcher
by PhD candidate
Extent of contribution
70
by PhD candidate (%)
CO-AUTHORS
Name Nature of Contribution
Michael P.N. Burrow Supervised the research and revised the paper
Harry T. Evdorides Revised the paper and methodology presented in the paper
Theunis F.P. Henning Supervised the research and revised the paper
Certification by Co-Authors
The undersigned hereby certify that:
the above statement correctly reflects the nature and extent of the PhD candidate’s contribution to this
work, and the nature of the contribution of each of the co-authors; and
in cases where the PhD candidate was the lead author of the work that the candidate wrote the text.
This form is to accompany the submission of any PhD that contains research reported in published or
unpublished co-authored work. Please include one copy of this form for each co-authored work.
Completed forms should be included in all copies of your thesis submitted for examination and library
deposit (including digital deposit), following your thesis Abstract.
Please indicate the chapter/section/pages of this thesis that are extracted from a co-authored work and give the title
and publication details or details of submission of the co-authored work.
"Using SVM to predict the probability of pavement failure"
This manuscript has been submitted to the Institute of Civil Engineers (ICE) Transport Journal in November 2012.
Nature of contribution
First author and researcher
by PhD candidate
Extent of contribution
70
by PhD candidate (%)
CO-AUTHORS
Name Nature of Contribution
Michael P.N. Burrow Supervised the research and revised the paper
Harry T. Evdorides Revised the paper and methodology presented in the paper
Certification by Co-Authors
The undersigned hereby certify that:
the above statement correctly reflects the nature and extent of the PhD candidate’s contribution to this
work, and the nature of the contribution of each of the co-authors; and
in cases where the PhD candidate was the lead author of the work that the candidate wrote the text.
INTRODUCTION
Maintenance planning is imperative for preserving the integrity of any road network and
heavily on a collaboration of historical network trends and inventory data, current condition
reports and field inspections, predicted performance of road pavements, and sound
Predicting the likelihood of road pavement failure and the associated causes is an important
aspect of maintenance planning. However, the causes of pavement failure are often complex
in nature and can be attributed to a number of factors including traffic loadings and
and strength, a weak subgrade, and poor construction practices. Such combinations, because
1
Chapter One: INTRODUCTION
of their complexity, are often overlooked in the diagnostic process despite the comprehensive
road reporting process. Furthermore, where road condition surveys rely on visual inspections
only, such as the case for many low volume roads that make up the majority of any country’s
road network, potential failures may not be recognised because, in these cases, visual
inspections commonly report only minor deterioration and disregard any obscured structural
problems. Although the data and knowledge associated with maintenance planning, is widely
New Zealand, alike to many countries, faces financial shortfalls in road maintenance funding
and currently its rate of increase in shortfall is greater than the rate of network growth and
inflation (Schlotjes et al., 2009). Given the financial pressures faced by road controlling
authorities (RCAs), the attitude towards asset management is shifting to a ‘do more for less’
approach, whereby decreasing the investment levels results in a reduced level of service
across the network. Under such an approach, the maintenance of roads of lower importance
and hierarchy in the network are often disregarded from the forecasted maintenance schedule
through deferring required maintenance and ignoring unexpected failures (SATCC, 2003).
from deferring maintenance and of ignoring unexpected failures, such that the asset manager
The case study, which forms much of the basis of this thesis, focuses on the greater New
Zealand road network, where the majority of the network consists of aged pavements
carrying low volumes of traffic (less than 10,000 vehicles per day). These pavements display
little visual deterioration over time, often failing unexpectedly (Arampamoorthy and Patrick,
2010; Bailey et al., 2006; Henning and Roux, 2008), and based on current practices, are
2
Chapter One: INTRODUCTION
generally excluded from the forecasted maintenance schedules (Schlotjes and Henning,
2012). Changes in the road environments, such as increasing weight and axle allowances of
heavy vehicles, increase the potential occurrence of unexpected failures across the network
(Schlotjes and Henning, 2012). From a practical viewpoint, unexpected failures have the
Currently, the forecasted maintenance schedules advise what maintenance treatments are
required, where such maintenance is needed, and when it should be implemented. Such
future state, expert opinion(s), and intervention criteria (Henning, 2008; Schlotjes et al.,
• Measure the shift in the failure probabilities and probability profiles over time;
assess this;
conditions, and
3
Chapter One: INTRODUCTION
A large amount of research has been carried out to develop models of road pavement
deterioration. Using these models, asset managers are able to forecast the expected life of
evaluation and assessment of the probability of road pavement failure, given its associated
environmental conditions, is lacking from pavement management systems (PMSs) (Salt et al.,
2010).
A number, and combination, of causes and physical attributes of pavements contribute to the
failure of road pavements. However, the interactions between these causes and environmental
conditions are lacking in the literature. These interactions are difficult to represent with a
conditions, and
• Errors associated with condition monitoring and variation included in the current
reporting processes.
Such methodologies have been successfully established for predicting failure probabilities of
other infrastructure assets, including water networks. Therefore, this research proposes a
methodology to:
pavement failure;
4
Chapter One: INTRODUCTION
techniques for evaluating the probability of road pavement failure from road
datasets, and
maintenance recommendations.
The aim of this research is to develop a framework that develops a prototype system capable
of quantifying the probability (likelihood) of road pavement failure and diagnosing the most
To achieve the above aim, this research has the following objectives:
failure;
2. Identify modelling techniques that can be used to predict the likelihood of failure
from road datasets and, from a comparative study, select the most appropriate for
3. Develop a prototype system based on the outcomes from the previous objectives,
which will assess the likelihood of failure for a given road site, and
5
Chapter One: INTRODUCTION
This research seeks to develop a robust and universal methodology that can be used to assess
the probability of road pavement failure, which is applicable to other pavement types and
road networks. However, the case study for this research focuses on New Zealand roads and,
as a result, Figure 1-1 identifies the road types addressed in this research. At least 95 % of the
New Zealand road network consists of flexible unbound pavements (Hayward, 2006), as it is
uneconomical for New Zealand to construct stronger, rigid pavements. The majority of these
roads are thinly designed based on the layer composition shown in Figure 1-2. The typical
design life for these pavement types is 25 years (Bailey et al., 2006), although
Arampamoorthy and Patrick (2010) reported the actual life of such pavements can be
6
Chapter One: INTRODUCTION
Managed by the New Zealand Transport Agency (NZTA), close to 100 % of the national
strategic road network of New Zealand, the State Highway network, is sealed. Only 61 % of
the local roads, managed by the Local Authorities, are sealed. Chipseal surfaces for this
research include single coat seals, two coat seals, void fill seals, slurry seals, racked in seals,
The geography of New Zealand has resulted in variable climatic conditions and road
environments, typically experiencing warm dry summer months and cold wet winters. The
variation in temperature between the seasons is not as severe as other countries, yet the
The case study of this research focuses on New Zealand road networks; therefore, the scope
stabilised;
7
Chapter One: INTRODUCTION
To achieve the objectives listed above, the thesis is structured as shown in Figure 1-3.
2. Chapter Three presents the framework followed in this research to develop the
prototype system. The generic nature of this chapter allows the transferability of
presents the causes of each failure mechanism with the use of failure charts. This
subsequent chapters;
8
Chapter One: INTRODUCTION
most appropriate technique to further develop for the specificities of the research
dataset;
5. The most appropriate modelling technique and developed failure charts are used
Chapter Seven. The case study demonstrates the network and project level
8. Chapter Nine concludes the thesis by presenting the accomplished work, the
9
10
Chapter Two
LITERATURE REVIEW
2.1 Introduction
model that is capable of predicting the probability of road pavement failure from road
datasets. The computational model will make use of engineering knowledge of pavement
failure to ensure that correct model independent variables are chosen. Therefore, to facilitate
this research, this chapter reviews current knowledge in the following areas:
modelling, and
11
Chapter Two: LITERATURE REVIEW
pavement failure.
The pertinent information from the review is then utilised to develop a theoretical framework
A large number of defects are associated with road pavement failure, depending on the
pavement types, the construction of road pavements, loading environments, and underlying
subgrade properties. However, since the New Zealand case study focuses on flexible,
unbound, granular pavements, this section reviews the predominant types of structural
compositions, namely rutting, cracking, and shear failure (Arampamoorthy and Patrick, 2010;
Creagh, 2005; Gribble and Patrick, 2008). Although roughness is an observed failure type, it
is predominantly regarded as a functional (and not structural) defect and, therefore, was not
considered further herein (Arampamoorthy and Patrick, 2010; Creagh, 2005; Robinson,
2008); instead Austroads (2007a) explains in detail the roughness failure mechanism, which
can be considered to be a measure of the combined effects of all of the other road surface
failure types, including rutting, cracking and shear failure types. While the subject of
Ruts appear in the wheelpaths of the pavement as longitudinal depressions resulting from
structural deformation of the layers below the wheel loadings (Papagiannakis and Masad,
12
Chapter Two: LITERATURE REVIEW
2008). The occurrence of rutting can indicate the deterioration of the structural integrity of
the pavement and pavement deficiency. If the ruts are wide and evenly-shaped, the
depressions are caused by weakness in the lower layers of the pavement. If the ruts are
narrow and sharply defined, the upper pavement layers are deficient in dissipating the
induced traffic loadings through the pavement (CSRA, 1992). Austroads (2007b) defines ruts
These depressions pose a safety hazard to the road user allowing water ponding to occur in
the surface ruts with the potential of aquaplaning and black ice1 hazards. Figure 2-1 shows
the appearance of rutting on a New Zealand road section. Recently, the noticeable increase in
accelerated rutting2 has become a problem on New Zealand roads (Henning et al., 2007).
Furthermore, the presence of fatigue cracking within the surface ruts is also noticeable on
New Zealand road networks, where it commonly occurs in pavements consisting of thin
1
Transparent ice on the wearing surface of the road resulting in a slick road surface, which poses an
inconspicuous hazard to drivers.
2
Defined by Henning et al. (2007) as the rapid deterioration phase of rut progression towards the end of its
design life, often once failure has been initiated.
13
Chapter Two: LITERATURE REVIEW
Rutting is accounted for in the Austroads pavement design using Equation 2-1 (Austroads,
2012), where the allowable number of Equivalent Standard Axles (ESA) loadings (N) is
calculated from the maximum compressive strain allowed at the top of the subgrade ().
Equation 2-1
=
9300
Henning (2008) adapted the World Bank Highway Development and Management Model
(HDM-IV) to replicate the performance of New Zealand road pavements, focussing on the
progression of rutting. In this study, Henning (2008) identified in the proposed rut
progression of rutting is swifter for thinner pavements than thicker pavements, under the
Cracking is defined as breaks in the integrity of the pavement surface indicating vertical
rupturing of the pavement, which may not necessarily extend through the entire thickness of
the surface or pavement layer(s) (Austroads, 2006). Different types of pavement cracking are
14
Chapter Two: LITERATURE REVIEW
• Edge cracking, is associated with the edge of the sealed pavement cracks, and
• Joint cracking, where the joints of different pavement subsections meet, and it
Surface cracking permits the ingress of water into the lower layers of the pavement, which
further exacerbates its progression (Austroads, 2006). As surface cracking progresses, the
cracks become interlinked and the surface seal disintegrates under continual traffic loading.
Fatigue cracking, shown in Figure 2-2, is the predominant cracking type associated with
structural failure on New Zealand road pavements (Henning, 2008), and largely appears in
the wheelpaths, indicating the association of this defect with the induced loading and strain
15
Chapter Two: LITERATURE REVIEW
Two crack initiation mechanisms are common in New Zealand. Figure 2-3a presents cracking
caused by large tensile strains at the bottom of the base layer where the cracks initiate form
beneath the surface. In this case, the design (fatigue) life of the base layer has been exceeded
by the current traffic loadings, resulting in a vertical divide in the base layer as the crack
further develops and propagates to the surface. Cracking as a result of insufficient pavement
support from below the base layer is shown in Figure 2-3b. The subsidence of the lower
pavement layers cause cracks to form in the top of the base layer in the wheelpath as the
material compresses under the wheel load (Austroads, 2006; CSRA, 1992; Henning, 2008).
(a) Cracking due to base layer performance (b) Cracking due to inadequate pavement support
Figure 2-3: Cracking mechanisms (Henning, 2008)
Fatigue cracking of an (unstabilised) asphalt layer is designed for using Equation 2-2
(Austroads, 2012), taking into account the volume of binder in the asphalt mix ( ), elastic
(Young’s) modulus of asphalt (E) and the maximum allowable compressive strain at the top
Equation 2-2
6918 ×
0.856 × + 1.08
ହ
=
.ଷ ×
16
Chapter Two: LITERATURE REVIEW
Henning (2008) investigated the initiation of cracking to be used as a detection point in the
forecasting of preventative road maintenance intervention for New Zealand. Henning (2008)
concluded that previously cracked pavements had a higher chance of cracking again, and
pavements with multiple surface layers were less stable than single surface layer pavements
due to significant flexing of the surface layers. The developed crack initiation model
Shear failure appears as potholes or deformations on the surface of the pavement resulting
from insufficient support of the underlying pavement structure. It is common for this type of
failure to occur on or near the edge of the pavement where it is particularly susceptible to
weakness from the ingress of seasonal moisture (Emery, 1992; Schlotjes et al., 2009; Thom,
2008). However, shear failures can also be seen across the entire pavement surface as a result
of material shear between the surface and pavement layers. Figure 2-4 illustrates a shear
To the road user, the ride quality diminishes with the presence of potholes and shoves on the
pavement surface, as well as posing a safety hazard. These pavement defects allow water to
enter the lower layers of the pavement, further adding to the variation of the seasonal
moisture zones located under the shoulders and near the edge of the pavement (see Figure 4-
2). New Zealand road networks are not normally susceptible to shear failures in periods of
17
Chapter Two: LITERATURE REVIEW
dry weather; however, after heavy periods of rainfall, it is common to see a large number of
There are limited design criteria surrounding shear failures due to the inherent behaviour of
material properties, the nature of the failure type, and poor performance in computational
models (Schlotjes et al., 2011; Schlotjes et al., 2012b). However, a number of recent studies
have been carried out to investigate the occurrence of shear failures on New Zealand roads.
Hussain et al. (2011) investigated the effect of varying moisture contents on common New
Zealand pavement aggregates, and concluded that the detrimental effect on the performance
of the pavement was a result of water ingress in the basecourse layer. Patrick (2009) reported
that shear failures were caused by the penetration of water through chipseal surfaces under
high traffic volumes. The study concluded that the permeability of the surface layer increases
especially in cases where water ponds on the surface after heavy rainfall periods, which
explains the increased number of reported potholes after heavy rainfall events.
18
Chapter Two: LITERATURE REVIEW
Flexible pavements are conventionally designed to fail by one failure mechanism (Austroads,
Zealand road networks, which creates difficulties in determining the primary cause(s) of
• Secondary effects, where a secondary failure type is the result of the primary
failure mechanism, such that rutting can cause cracking and vice versa.
Failure Mode and Effect Analysis (FMEA) and Fault Tree Analysis (FTA) techniques are
widely used to recognise and understand the causes of future occurrences of failure of a
system (Isa et al., 2005; Lindhe et al., 2009; Patev et al., 2005; Pickard et al., 2005; Xiao et
al., 2011).
FMEA, first developed in the 1960’s, is an analytical tool utilised in reliability analysis
studies, such as the management of engineering systems, to assess the reliability and safety of
system processes (Seyed-Hosseini et al., 2006). As a result of the analysis, the possible
failure cause affecting the system’s functionality and the consequences of such failure are
identified to minimise or eliminate failure in systems. It aims to answer the questions of:
‘What might go wrong?’, ‘What could cause it to go wrong?’, and ‘What would the effect of it
be?’ (Seyed-Hosseini et al., 2006). The consideration of subjective and expert knowledge
results in a weighted ranking system that assigns a priority (risk) value to the possibility of a
19
Chapter Two: LITERATURE REVIEW
failure event. From this, the overall impact of an event can be calculated using Equation 2-3
below.
FMEA has been successfully employed in a number of fields to combine the effect of
multiple failure modes (Pickard et al., 2005; Xiao et al., 2011), prioritise failures based on an
adjusted risk priority number (Seyed-Hosseini et al., 2006), and with the use of functional
models automate the FMEA process of complex engineering systems (Hawkins and
Woollons, 1998). Further development of the approach includes using fuzzy logic, causal
reasoning, and Boolean representation of the inputs and outputs (Bell et al., 1992; Wang et
al., 1995; Xu et al., 2002) to improve the assessment of the known and unknown variables.
The causes of failure can be presented in a graphical manner using FTA, such that
failure (Patev et al., 2005; Yuhua and Datao, 2005) and presented as failure paths. The
manner to that of FMEA. This approach includes logic AND gates and OR gates, allowing for
multiple causes of failure to be included in the binary assessment of the system and
increasing the number of scenarios the approach can replicate (Lindhe et al., 2009). However,
unlike the FMEA, no consequence of the event is included in the FTA approach. Despite this,
the benefit of this approach includes the representation of multiple causes, and the
interactions between such causes, of failure (Patev et al., 2005; Schlotjes et al., 2012b).
FTA has been adopted in manufacturing, chemical, and other engineering fields. Such studies
include using fault trees to estimate the failure probability of gas pipelines (Yuhua and Datao,
20
Chapter Two: LITERATURE REVIEW
2005), building binary decision diagrams from fault trees to calculate event probability (Reay
and Andrews, 2002; Remenyte-Prescott and Andrews, 2008), and risk probabilities of
drinking water systems (Lindhe et al., 2009). Development work to improve the FTA
approach includes incorporating human logic into the system (Ortmeier and Schellhorn,
2007), modelling sequential failure logic in fault trees (Long et al., 2000), and using fuzzy
However, although there appears to be little information in the literature about the use of the
FMEA and FTA approaches in pavement performance modelling, the concepts of each of
these approaches are applicable to the analysis of road pavement failure. Consequently, their
computational model for predicting road pavement failure is explored further in Chapters
probabilistic (Haas et al., 1994; Lytton, 1987; Martin, 2008). Figure 2-5 graphically defines
21
Chapter Two: LITERATURE REVIEW
Deterministic: The outputs of the model are determined through known statistical
Randomness does not play a role in the future model predictions. Therefore, for a
given initial state, the model will always produce the same result, such that the exact
output is known once the system begins. Such models yield future values of the
forecasted parameters, which in practice have been used widely for both network and
where the model predicts the variability and probability of the dependent variable.
The outputs of the model are determined through variables described by probability
distributions as opposed to unique (discrete) values. Therefore, the model can predict
the likelihood of an event, but the timing or the occurrence of the event is unknown.
The mean of the probability distribution defines the most likely output estimate. As
these model types demand computing resources, they are predominantly used at a
22
Chapter Two: LITERATURE REVIEW
below.
model. Fundamental knowledge, first principles and laws are referred to and used in
relationships between the inputs and output(s) and defining the model parameters.
These methods are useful when the behaviour of pavement failure is not fully
understood. However, the transferability of the developed models outside of the data
using observational data. These models can be used outside of the data domain, only
Linear: A linear function takes the form of Equation 2-4, resulting in a straight-line
relationship between the dependent (output) and independent variables (inputs). This
multiplication) properties (Bapat, 2000; Pruim, 2010), inferred by the constants c and
m in Equation 2-4. Given the complexity of pavement performance data (refer Section
23
Chapter Two: LITERATURE REVIEW
1.2) (Reigle, 2000), linear model forms are not expected to be successful in accurately
= "# + !
Equation 2-4
independent variables and further does not satisfy the associated properties of linear
Model outputs can take the form of classification or regression, further explained
below.
Classification: The model outputs take a discrete class label, such as a binary value
{0, 1}, or a group membership from a list of categorical expressions, such as material
types {rock, sand, clay}. Pavement failure is represented in this research as a binary
class label, where a failed pavement is assigned a one (1) and a sound pavement is
Regression: The values of the outputs take continuous values, opposed to discrete
values alike above, where the dependent variable of a regression model is numerical.
Early pavement deterioration models employed simple regression analyses; but now
24
Chapter Two: LITERATURE REVIEW
Performance models play an integral part in the management of infrastructure assets where
forecasting future maintenance requirements heavily rely on the predictions from such
models. Numerous studies have employed both deterministic and probabilistic approaches in
the development of performance models, focussing on two phases of the asset’s life:
• Asset failure: The end of life of the asset, when replacement, reactive or remedial
The success in the development of deterioration curves has increased the popularity of
predictive methods for infrastructure assets, including bridges (Basheer et al., 1996; Lounis
and Madanat, 2002; Madanat et al., 1997; Pan et al., 2009), water, wastewater and
stormwater networks (Abraham et al., 1998; Ariaratnam et al., 2001; Baik et al., 2006;
Kleiner et al., 2001; Osman and Bainbridge, 2011; Wirahadikusumah et al., 2001; Younis and
Knight, 2010), and others (Morcous et al., 2002; Sharabah et al., 2006).
Both stochastic and deterministic approaches have been explored in pavement deterioration
studies, and include Markovian processes (Wang et al., 1994; Yang et al., 2005; Yang et al.,
2006), Bayesian probabilities (Amador-Jiménez and Mrawira, 2011; Gao et al., 2012), and
regression analysis to relate the independent and dependent variables (Arambula et al., 2011;
Ben-Akiva and Gopinath, 1995; Gupta et al., 1997; Isa et al., 2005; Manik et al., 2007;
Martin, 2008; Park et al., 2001; Prozzi and Madanat, 2002; Sadek et al., 1996). Markovian
processes exhibit stochastic behaviours often used in Markov chains, where the next state in
25
Chapter Two: LITERATURE REVIEW
the chain is solely based on the current state and not the preceding events in the chain
(Austroads, 2009b; Wang et al., 1994). Bayesian probabilities enable reasoning to be inferred
in models mathematically through the use of probabilities and prior evidence (Amador-
Henning (2008) investigated the use of the generic mechanistic-empirical HDM models for
New Zealand road pavements and identified limitations with the linear regression approach
taken in the development of these models. Henning (2008) found that the generic HDM
models were not able to accurately predict pavement deterioration and instead developed
logistic regression models (see Section 2.6.1) for predicting crack initiation and accelerated
rutting, using the development of the HDM models as a benchmark. The proposed logistic
regression models considered additional influential factors, such as pavement thickness when
calculating the probability of accelerated rutting, which was previously omitted from the
modelling. However, failure models developed for other infrastructure assets were discussed
in the literature, such as for power systems (Halilčević et al., 2011; Hu et al., 2011; Rau et al.,
1977), water networks (Anghel, 2009; Zhong et al., 2011), mechanical engineering systems
(Pickard et al., 2005) and others (Cai, 1996). Previously, stochastic approaches were
employed to generate probabilities of failure, such as failure or fault trees, fuzzy logic, and
probability theory (Cai, 1996; Hu et al., 2011; Pickard et al., 2005). However, with the recent
26
Chapter Two: LITERATURE REVIEW
Anghel (2009) investigated the performance of classification approaches to predict the failure
probability of pipelines. The predicted output was successfully calculated with a binary
support vector machine classifier. Included in the methodology was a definition of the failure
modes in the model inputs. An assessment of the performance of the independent variables
influenced the model parameters, including the choice of kernel transformation (see Section
2.6.3), such that industry knowledge is inferred in the computational model. The results from
this study were comparative to traditional stochastic methods used in predicting pipeline
failures.
A variety of complex factors, some uncertain, play a role in the failure of dams. Zhong et al.
(2011) employed stochastic methods to calculate the probability of dam failure risk, based on
the individual probabilities from the three selected failure types. Probability theory was
employed in combining these into an overall failure assessment, assuming the factors
contributing to dam failure were independent. To employ such theory, the authors discussed
the importance of identifying the factor relationships prior to calculating the overall failure
probability.
Cai (1996) defines failure in terms of a predetermined failure criterion; either the system has
failed or not, and represents the outcome in a binary format {0, 1} where failure is
represented by one (1) and zero (0) represents non-failure. This study investigated the use of
fuzzy logic in safety, reliability, and risk assessments of system failure to improve other
model formats, such as deterministic or probabilistic. Employing fuzzy logic in such models
27
Chapter Two: LITERATURE REVIEW
(see Section 2.6.7) accounted for the fuzzy undefined states of variables in engineering
systems.
In the transportation sector, Arampamoorthy and Patrick (2010) based the predictions of
pavement failure probability solely on the historical trends of four road networks, using
regression analysis techniques, and deterministic design criteria. This simplified approach did
not identify the interactions between factors causing failure; instead the study included only
pavement strength, composition, or subgrade properties. A further limitation of this study was
Classification techniques, both discriminative and generative methods, have been employed
by other researchers to predict an event, such that the predicted output takes a discrete value.
and separate the classes of data using a decision boundary. On the other hand, generative
classifiers define an event from the observed data using a probabilistic distribution of each
variable. Therefore, the latter are employed to predict a classification from a probability
density function, as opposed to the former that predicts to which class the new observation
The literature reports better performance from discriminative classifiers than generative
methods. However, excluding generative classifiers from the literature search preconceives
28
Chapter Two: LITERATURE REVIEW
the notion that one technique is superior over another. Therefore, this section investigates
• Logistic regression;
• Neural networks;
• k-nearest neighbours.
Logistic regression separates the data into classes with the use of a linear decision boundary
(Eastaugh et al., 1997). The popularity of this technique is attributed to the transparency of
the learning algorithm and the identification of causal relationships within the data (Tu,
1996). Despite the apparent simplicity of the technique, logistic regression has performed
variables; although this technique reported a better performance on smaller datasets (Lim et
al., 2000). The limited computational power required and quick training time makes this
technique attractive to researchers and practitioners; however, the linearity of the decision
problems (Lim et al., 2000; Razi and Athappilly, 2005). However, logistic regression has
been applied to road pavement data, leading to the successful employment in the
development of deterioration models for New Zealand road networks as described earlier in
Perlich et al. (2003) investigated the performance of logistic regression models in classifying
binary outcomes on several datasets, ranging from 1000 to several million datapoints, in
29
Chapter Two: LITERATURE REVIEW
various subject areas. As reported in this study, the logistic regression technique performed
well on smaller datasets; however, other techniques tested, such as decision trees,
outperformed logistic regression on larger datasets due to the clear class separations included
in the larger datasets. The authors were apprehensive on drawing conclusions on superior
Peng et al. (2002) found logistic regression to be effective in predicting binary outcomes in
regression and reported on the success of the technique across a wide variety of datasets. In
generally outperformed when continuous data was used. Austin et al. (2010) also reported on
the superior performance of logistic regression, against decision trees, in predicting mortality
of hospitalised patients.
Many studies have compared the performance of logistic regression models to that of neural
1997; Eftekhar et al., 2005; Saghafi et al., 2009; Tu, 1996). In most cases, neural networks
outperformed logistic regression, as the latter employs a linear decision boundary (Eastaugh
et al., 1997). However, Eftekhar et al. (2005) reported a higher accuracy for logistic
regression, although the authors employed alternative performance measures given the
erroneous and bias performance assessments associated with accuracy (Ben-David, 2008;
30
Chapter Two: LITERATURE REVIEW
Neural networks, also known as multi-layer perceptrons, have become popular in machine
learning applications, given their ability to determine non-linear relationships within the data,
even in the presence of noisy and incomplete data (Jagielska et al., 1999) or without knowing
the relationships between the independent and dependent variables (Razi and Athappilly,
resources, and is described as a ‘black box’ approach (Tu, 1996). Neural networks are known
to locate local minima3, yet researchers are not adverse to employing neural networks given
the infinitesimal difference between the local minima and global minimum4 (Eastaugh et al.,
1997). The complexity, in particular the non-linear decision boundary, of this technique has
techniques (Dreiseitl and Ohno-Machado, 2002; Eastaugh et al., 1997; Saghafi et al., 2009;
Tu, 1996).
Kaseko and Ritchie (1993) employed neural networks to detect and classify pavement
cracking. Using images of the road pavement surface, neural networks were employed to
distinguish if the pavement was cracked and, furthermore, the type of cracking present on the
pavement surface. Initial results showed the multi-layer feed-forward neural network model
had difficulty (67 % accuracy) in distinguishing between two types of alligator cracking, yet
the model performed well for transverse, longitudinal, and block cracking. However, when
distinguishing only between the types of alligator cracking, the detection accuracy of the
3
A possible solution to a mathematical problem, which is located in the neighbourhood of the exact minimum
solution, and is often referred to as a relative solution.
4
The unique solution to a mathematical problem that is associated with the smallest error, and is often referred
to as the absolute solution as there is only one optimal solution to the problem.
31
Chapter Two: LITERATURE REVIEW
classify soil types using data from cone penetration tests. In the cases of binary classification,
the neural network model accurately classified sandy soil samples but, in the case of
In the power industry, an index describing the transient stability (a binary output) of power
systems was developed using neural networks (Tso et al., 1998). Using a training dataset of
200 samples, including a range of fault locations and loading conditions, the model structure
included one hidden layer and employed 11 input variables. On testing the applicability of the
trained classifier, only 3 % of the test samples from an independent evaluation dataset were
misclassified. The authors recognised the transferability of the neural network model to other
The medical industry has employed neural networks as an alternative classification method to
linear approaches such as logistic regression; although both of these techniques remain
popular (Dreiseitl and Ohno-Machado, 2002). Eastaugh et al. (1997) investigated the use of
neural networks in predicting high risk pregnancies. Although this study identified the
shortfall of neural networks converging to a local minimum, the results showed a slightly
higher accuracy from the neural network model than logistic regression. However, Dreiseitl
approaches when used in binary classification tasks and concluded no single algorithm
outperformed another. Instead, logistic regression has been popular in the medical industry
given its interpretability and user-friendliness, whereas the complexity from the non-linear
decision boundary has led to the successful employment and popularity of neural networks.
32
Chapter Two: LITERATURE REVIEW
Recent advances in machine learning algorithms have resulted in the suitability of support
vector machines as a classifier (Cortes and Vapnik, 1995). Despite the similarities with neural
networks, support vector machines locate the global minimum and require less computational
resources than neural networks resulting in its increased popularity. With the addition of
kernel functions5, engineering knowledge can be inferred in these non-linear models (Noble,
2004) by an appropriate choice of the transformation function (for class separation), given the
Gualtieri et al. (1999) reported the early success of support vector machines as a binary
classifier, while the algorithm was in the early stages of development, and suggested the
networks, this technique is less transparent than linear models but is more interpretable than
neural networks. Support vector machines require fewer input variables and a less complex
network structure than neural networks, for generally better classification capabilities (Pal,
2006).
Bhattacharya and Solomatine (2006) reported that support vector machines outperformed
neural networks in classifying non-sandy soil types, yet the overall accuracy of the two
techniques differed by only 0.2 %. Pal (2006) further recommended support vector machines
utilising the radial basis function kernel (see Table 5-2) instead of neural networks for soil
classification, given the more accurate results produced and less training time required. The
generalisation capabilities of this technique and the transferability of the technique to other
5
Functions that transform the data into a higher level space where a simpler classification separation is capable.
Knowledge of the data can be included in the model, based on the choice of kernel transformation.
33
Chapter Two: LITERATURE REVIEW
datasets were discussed by Pal (2006), given the improved performance of the developed
support vector machines model when tested with a dataset unfamiliar to the model.
The similar nature of support vector machines to neural networks has seen these models
developed for the medical industry. For example, Gil and Johnsson (2011) reported
accuracies greater than 90 % (see Table 2-1) in diagnosing medical dysfunctions using
support vector machines with a five-fold cross-validation sampling method applied. The low
precision of the results for the urological database was attributed to the quality of the data in
comparison to the other two studies. The authors discussed the complications with binary
classification, particularly when the data reflects skewed distributions, which could be
reflected in the results; such distributions are evident in road pavement data.
decision boundary; instead the algorithm splits the data optimally at each node, accordingly
to the classes of the data (Dreiseitl and Ohno-Machado, 2002), and is a favoured choice for
easily separable classes (Perlich et al., 2003). The interpretability of decision trees lends itself
to studies in various research domains despite the performance of the technique lacking
competitiveness against other modelling approaches (Chen et al., 2004). This technique has
34
Chapter Two: LITERATURE REVIEW
performed adequately with smaller datasets and models developed using this algorithm are
Chen et al. (2004) presented an automated system for the failure diagnosis of internet faults.
This study employed decision trees to diagnose the cause of identified faults achieving a
success rate of 93 %. Although the authors recognise the availability of other classifiers,
generally with better failure prediction performance, they found that the interpretability and
Bhattacharya and Solomatine (2006) reported comparative results for decision trees against
more sophisticated models, such as neural networks and support vector machines. In fact, in
this study the overall performance of the decision trees was similar to that of neural networks
in classifying soil types. Razi and Athappilly (2005) compared the performance of decision
trees to neural networks and linear regression models, such as logistic regression, in
predicting the health of an individual. This study concluded that decision trees performed
well against neural networks and both of these techniques were in fact substantially better
performers than linear regression techniques. This was further supported by additional studies
carried out, such that Perlich et al. (2003) reported a superior performance of decision trees
A number of decision trees can be collated to create random forests through the method of
bagging, where equal weightings are assigned to individual decision trees assuming each tree
is equally important to the overall result (Roiger and Geatz, 2003). Through this process, the
benefits and attributes of decision trees are inherited and furthermore, as a result of bagging,
random forests improve on the stability of individual tree models. As an alternative classifier,
random forests are becoming popular given the model simplicity, inherited from decision
35
Chapter Two: LITERATURE REVIEW
trees, in comparison to support vector machines and neural networks (Verikas et al., 2011),
although the interpretability of decision trees is not conveyed in random forests. Random
feature selection is also employed throughout the creation of a random forest (Prinzie and
Chandra et al. (2009) reported on a comparative study between random forests, decision
trees, support vector machines, a feed-forward neural network, and logistic regression models
in predicting the failure of internet (dotcom) companies, given historical data. In this study,
random forests outperformed the other four methods; however, the authors found that using
two or more techniques to produce a series of classifiers exceeded the performance of any
individual model.
Verikas et al. (2011) also investigated the performance of random forests against the
performance of other popular techniques, such as support vector machines, neural networks,
and decision trees. A variety of datasets were used to evaluate the classification accuracy of
the methods, ranging in a number of classes, including binary, and sizes. Although this
technique demonstrated superiority on large scale comparisons over the other trialled
transparent modelling processes, of the technique (Rogers and Girolami, 2012). Only one
parameter (k) is adjusted in the model, opposed to neural networks, for example, which
requires the user to define many model parameters. However, the simplicity of the technique
often leads to performance deficiency and inappropriate use in complex problems. On larger
datasets, this technique can take a long time to train and becomes computationally expensive.
36
Chapter Two: LITERATURE REVIEW
As the user defines the distance between the nearest neighbours, this model can misrepresent
the problem when the relationships between the data are unknown or difficult to define
from this technique than decision trees, logistic regression, and naïve Bayes. However, other
discriminative classifiers, such as neural networks and support vector machines, were found
to outperform this technique. Dreiseitl and Ohno-Machado (2002) reported on the poor
performance of k-nearest neighbours favouring neural networks and logistic regression over
this technique, especially on high-dimensional data, despite the simple and easy model
patterns and relative importance criteria of dimensional data into the model, to improve the
Despite the reported success of discriminative classifiers over generative approaches (Long et
al., 2007), it was necessary herein to investigate the following generative classifiers:
• Naïve Bayes;
Table 2-2 summarises the advantages and disadvantages of the above techniques.
37
Chapter Two: LITERATURE REVIEW
38
Chapter Two: LITERATURE REVIEW
boosting, and bagging methods. While such approaches were considered to be beyond the
scope of this research, a brief description of each is given below for completeness.
Rough sets, fuzzy logic, and genetic algorithms, are often used to enhance the performance of
problems (Hu, 2010; Jagielska et al., 1999; Örkcü and Bal, 2011). Rough sets can account for
uncertainty based on exact rules and definitions of the upper and lower boundaries attaining
to the specific data (Jagielska et al., 1999). Fuzzy logic accounts for uncertainty and
imprecisions in the data, based on approximation reasoning, and characterises the ‘grey’ area
between data classes (Pan et al., 2011; Yuhua and Datao, 2005). Studies employ genetic
algorithms to solve engineering optimisation problems (Bernard et al., 1998; Morcous and
Lounis, 2005).
Boosting refers to the methodology of compounding a stronger model from a set of weaker
performing classifiers, in a successive order, based on the results from the individual
antecedent models. The later models concentrate on correctly classifying the datapoints that
have been incorrectly classified in the earlier phases of the model sequence. On completion
of the training phase, a weighting factor is assigned to each model in the sequence, based on
the performance of the individual model on the training data (Roiger and Geatz, 2003).
models involved assuming equal importance of each model to the overall result (Roiger and
6
A collection of different techniques used concurrently as one model.
39
Chapter Two: LITERATURE REVIEW
Geatz, 2003). Random forests are an example of this methodology and, since only multiple
decision trees (one model type) are combined, are considered in this research (refer Section
2.6.4 above). Composite bagging methods compound various and assorted modelling
Table 2-3 summarises the key points from the literature review on classification techniques.
40
Chapter Two: LITERATURE REVIEW
41
Chapter Two: LITERATURE REVIEW
In addition to the above, key elements from the literature utilising these binary classifiers are
summarised below:
• There is not one superior technique for all domains and datasets (Perlich et al.,
2003). The performance of the classifier greatly depends on the specific dataset;
therefore, an assessment of the classifiers against each other using the research
dataset will be required to identify the classifier most suitable for the task at hand;
other performance measures should be used to account for the biasness associated
2011);
• Poor quality and incomplete datasets result in low precision and accuracy of the
• When the data class distributions are unequal, problems with binary classification
occur. During the training phase, a model will always try to minimise the overall
model that will always predict the larger of the two binary classes (Prinzie and
Van den Poel, 2008). Given that unequal class distributions are a reality in
network datasets7, a weighting factor should be applied in the training phase of the
7
Road failure, as a result of preventative maintenance, is less frequent on the network than the number of sound
pavements. Therefore, this is reported and represented in the data.
42
Chapter Two: LITERATURE REVIEW
models to ensure equal preconception of the two classes (failure and non-failure)
• The majority of the studies, across all classification techniques, tested the model
one-third of the entire dataset. Larger training datasets will result in a better
generalisation of the model, whereas a larger testing dataset will present a better
• Son et al. (2012) identified the importance of identifying relationships, both linear
Through the review of the literature, the approaches considered by other researchers were
scrutinised. Firstly, from a fundamental perspective, FMEA and FTA have been used to
43
Chapter Two: LITERATURE REVIEW
represent knowledge of failure in computational models. FTA has the ability to represent
failure as a binary outcome and can represent multiple causes of failure and interactions
between failure factors. Therefore, this research adopted the FTA approach to infer
engineering knowledge into the computational model, further described in Chapter Four.
Researchers have developed deterioration models for road pavements using a variety of
methods, including both deterministic and stochastic approaches. Deterioration models have
Although limited studies focussed on predicting the probability of pavement failure, other
industries have successfully developed methodologies to address such aims. One study
as listed in Table 2-3. Classification techniques have been used successfully in a number of
areas, ranging from medical diagnosis, power systems, geomechanics, and other engineering
fields. From the literature, neural networks and logistic regression emerged as popular
choices of binary classifiers; however, Table 2-3 describes the benefits for the employment of
the other classification techniques. Since the performance of these techniques was shown to
be greatly dependent on the datasets used in the studies, a comparative study (see Chapter
techniques identified in this chapter for the task of predicting the failure of road pavements
44
Chapter Three
METHODOLOGY
3.1 Introduction
The literature review demonstrated there has been little research to develop robust
probability of road pavements. To address this, this chapter presents the systematic approach
followed to quantify the probability of pavement failure, addressing the aim of the research.
Although this research concentrates on the structural failure of chipsealed flexible road
pavements, the approach outlined can be adapted for other road pavement types and failure
mechanisms.
45
Chapter Three: METHODOLOGY
Failure charts for each of the three failure mechanisms, outlined in the scope of this thesis
(refer Section 1.3.3), were developed to form the diagnostic element of the system and were a
direct output from the first research objective. The computational model formed the
probability element of the system and was developed using a technique selected from an
analysis, using objective criteria, of a number of techniques described in the literature. With
this and the failure charts, the system can determine the probability of failure and identify the
most probable failure path, deemed to be the critical failure path in this research.
associated with road pavement failure, and utilising this to develop failure charts;
• Investigating the performance of the selected techniques to accept one as the most
46
Chapter Three: METHODOLOGY
Identify UNDERSTANDING
Failure Types ROAD PAVEMENT
Understanding Road FAILURE
Data Exploration – Descriptive Pavement Failure
Road Condition Statistics of
Datasets Parameters
Identify
Influential
Parameters
Criteria for
Selecting Model
Forms
Literature Review
(Modelling
Techniques) TRIAL MODEL
SELECTION
Select Model(s):
Identify Models for Further
Consideration
Develop the
Research
Prototypes of
Dataset
Models
Performance
Measures FINAL MODEL
TECHNIQUE
Assess Each SELECTION
Model Prototype
Criteria for “Best”
Model
Select
“Best” Model
Understanding Road
Pavement Failure
Performance
Evaluation of
PROTOTYPE
SYSTEM
47
Chapter Three: METHODOLOGY
datasets, comprising of descriptive statistics and correlation analyses of the available road
datasets including a typical RCA network dataset, and a review of the literature identified the
most important factors that contribute to road pavement failure. Incorporating engineering
knowledge into the modelling process resulted in a more accurate and knowledgeable model
(Schlotjes et al., 2011; Schlotjes et al., 2012a). In order to facilitate the conceptual design of
the research and infer engineering knowledge within the computational model, failure charts
were used to present the key factors and combinations contributing to road pavement failure.
The approach to the development of the failure charts was based on FTA (refer Chapter
Two), which recognises the importance of identifying the causes of failure and understanding
the interactions between the possible failure causes (Patev et al., 2005). The graphical format
of the failure charts was found to be a convenient method of presenting the complex
Figure 3-2 outlines the process of developing the failure charts. The first task was to define
failure for each of the failure mechanisms (e.g. rutting, cracking, and shear). The definition of
failure for each mechanism assisted in identifying the causes of failure, as well as the
interactions between these causes. From this, the combinations of multiple failure causes
represent the failure paths for each failure mechanism. Chapter Four further describes how
the failure charts were developed using information from the literature and expert opinion,
together with specific methodology for the development of an understanding for flexible
pavement failure.
48
Chapter Three: METHODOLOGY
Chapter Two has reviewed the use of computational techniques in a number of fields,
including the transportation sector, which may be suitable for modelling road pavement
performance. As pavement failure is defined in the research dataset as a binary output {0, 1}
(see Section 3.3.1.7), the research problem becomes a binary classification problem. Table 2-
3 identified a number of classification techniques for further consideration and the suitable
the literature;
2. Road pavement datasets are potentially extensive and may consist of large
amounts of data relating to many variables, thus the modelling technique needs to
road pavement data (Reigle, 2000). Complex problems are better represented with
this study, with the provision of using linear methods with caution;
4. The time required to run the selected model should be kept to a minimum to
5. The training phase of the model should be simple, easily inferred, and based on
49
Chapter Three: METHODOLOGY
6. The overall process should be simple and easily understood by asset managers and
processes, and
The detailed investigation of the suitability of the techniques is discussed in Chapter Five.
The selected techniques were compared using a research dataset (see Section 3.3) to select
the most appropriate technique for the development of the prototype system, as the literature
reviewed in Chapter Two concluded there is not one superior technique for all domains and
datasets. This was achieved by assessing the output from each technique according to four
2007; Parker, 2011). However, this research extended the assessment criteria to include
interpretative qualities and usability characteristics of the techniques in reference to the future
2. The speed of the model, simplicity of use, and the overall process is easily
interpreted, and
Chapter Five further describes and discusses the performance assessment of the five trialled
modelling techniques.
50
Chapter Three: METHODOLOGY
As described earlier in Section 3.1.1, the conceptual design of the system consisted of two
elements:
1. Failure charts, from the knowledge of failure, to diagnose the cause(s) of failure,
and
Therefore, the development of such a system consisted of failure charts and a computational
model, as described in Sections 3.2.1, 3.2.2, and 3.2.3. The prototype system was developed
from the failure paths presented in the failure charts, with each path being modelled by the
computational technique. The resultant system predicted the probability of failure of a road
section for each failure mechanism, which was calculated as the maximum (most probable)
probability of each failure path, per failure mechanism. The statistical package R8 (Everitt
and Hothorn, 2006) constructed the prototype system, as well as the classification techniques
trialled in the comparative study; although the purpose of this research was not to assess this
statistical package.
pavement section taking into account all failure types. The first was a simple approach that
assumed that each failure mechanism acts independently, following the approach adopted in
ிூோா = "# $
%,
&, … , ()'
Equation 3-1
ℎ: =
, =
, =
8
A number of packages are readily available to be used for the task at hand; however, given the access to a large
library of model files, R is capable of modelling each of the techniques outlined in Chapter Two.
51
Chapter Three: METHODOLOGY
However, a more complex model considered to be more realistic was also developed that
considered the interdependence of each failure type using the ‘Additive Law of Probability’
(probability unions), by which multiple failure mechanisms can be combined (Ayyub and
Equation 3-2
∪ ∪ = + + − ∩ − ∩ − ∩ + ( ∩ ∩ )
=
, =
, =
= + + − × − × − × + ()
× () × ()
(ℎ: 0 ≤
∪ ∪ ≤ 1
=
, =
, =
Equation 3-2 calculates the probability of either failure mechanism occurring using the
maximum probabilities from each failure mechanism, as described previously (refer Section
3.2.4). The description of the development of the system and interactions between the failure
prototype system.
The prototype system was tested using an independent dataset which had not been used to
train the computational model. The test consisted of comparing the probabilities of failure
computed using the prototype system with the number of actual failed sections in the dataset.
52
Chapter Three: METHODOLOGY
Chapter Seven presents a detailed analysis of the prototype system testing, along with a
The New Zealand Long-Term Pavement Performance (LTPP) programme was established in
2000 and consists of a number of sites where data is gathered from New Zealand road
networks. This research used two datasets compiled from the LTPP database; one of the road
sites on the State Highway network, and the other from sites across the Local Authority road
networks. The primary difference between the two datasets is the external environment of the
The benefit of the LTPP programme is the frequency of data collection and the high level of
detail of the data, which has resulted in a sufficient amount of data that may be considered
appropriate for such research studies. Traffic data, inventory data, condition data, detailed site
inspections, and maintenance reports are included in the LTPP programme database (Henning
et al., 2004). Each site is surveyed annually to record the condition of the pavement using
Figure 3-3 illustrates the layout of a typical LTPP inspection site. Data is collected from the
first 300 metres of each site with the length of the site divided into six 50 metre sub-sections.
53
Chapter Three: METHODOLOGY
From the State Highway LTPP dataset the following data types were obtained:
• Traffic;
• Pavement composition;
• Pavement strength;
• Environment;
• Subgrade sensitivity;
• Failure data.
The dataset contains the Annual Average Daily Traffic (AADT) and Percentage of Heavy
Commercial Vehicles (HCV). However, it was felt necessary to convert these into an
54
Chapter Three: METHODOLOGY
(Cum_ESA), which takes into account the impact of heavy vehicles on road pavements over
time. This was achieved using Equation 3-3 (Austroads, 2012) as follows:
Equation 3-3
"_)% = * ா × +ி × 365
%%,- × × , × .
=* × +ி × 365
100
= ! "#
ℎ
!!
ℎ
1
=
ℎ
=
! $
1 + 0.01
' = 'ℎ ݂ܽܿ= ݎݐ
0.01
The State Highway LTPP dataset used historical traffic data to calculate the annual traffic
growth rate r and the breakdown of the HCV classes to calculate the ESA per heavy vehicle.
Furthermore, the dataset obtained two Cum_ESA parameters from using both the ages of the
surface layer and the base layer in the calculation of cumulative number of ESAs.
composition data of the LTPP pavements. The information available included the material
properties of the pavement layers (surface layer, base layer, and sub-base layer, if applicable),
55
Chapter Three: METHODOLOGY
the layer ages, layer thicknesses, number of lanes, and pavement widths; however, some of
The State Highway LTPP dataset contained four factors associated with the strength of a road
pavement section. The modified structural number (SNP) is back-analysed from deflection
readings and its average value, along with a binary strength parameter which indicates
whether the pavement is weak or strong, was utilised in this research dataset. A pavement
was considered strong in this research if (i) SNP ≥ 2.5, or (ii) the thickness of a single layer
exceeded 250 mm. Individual deflections from the falling weight deflectometer (FWD) were
available, from which the following parameters were calculated for comparative measures of
0
Equation 3-4
/ =
ଶ
,ଶ
2 × , × 11 − 2, 3
56
Chapter Three: METHODOLOGY
Finally, a concurrent New Zealand study developed structural indices algorithms which, by
failure mode (rutting, flexure, shear, and roughness) (Salt et al., 2010). Each LTPP site was
analysed using the structural indices and were included in the State Highway LTPP dataset.
The LTPP programme obtains rainfall data from the National Institute of Water and
Atmospheric Research (NIWA) and, accordingly, the State Highway LTPP dataset contains
the annual rainfall values (in millimetres). Water entering the pavement once the surface seal
has cracked is a key factor in the deterioration of the pavement structure, which is not a factor
included in the LTPP programme. Therefore, it was felt to develop a new variable that is a
measure of the cumulative rainfall once the pavement is cracked (CumRain_ifCracked), using
A design matrix included in the LTPP analysis (Henning, 2008) categorises the pavement
sites into sensitivity classes based on the properties of the underlying subgrades. This
accounts for the subgrade material types, drainage, and environmental conditions. This factor
classifies the sites into low, medium, and high sensitivity zones.
This research defined three failure mechanisms that have been found to be prevalent on low
volume roads in New Zealand, namely rutting, cracking, and shear (Arampamoorthy and
57
Chapter Three: METHODOLOGY
Rutting Data:
The rutting depths on the LTPP pavement sites were evaluated using a transverse profile
beam. For each wheelpath, the rutting depth (in millimetres) is calculated by averaging two
measurements taken at the same point on the road surface. Following this, the method is
repeated every 10 metres from the start (0) point, resulting in an average rut depth for each
wheelpath every 10 metres. Along each 50 metre subsection (1 through to 6) (refer Figure 3-
3), the rut depths are further averaged for a resultant rut depth per wheelpath per subsection.
A combination of the results from the two wheelpaths provides an overall estimate of the rut
depth for the lane as a whole. For completion, the State Highway LTPP dataset included the
rut depths for the both of the wheelpaths and the lane (increasing or decreasing). Further
manipulation of the rut depths produced a rut rate (per year) for each subsection.
Cracking Data:
The total amount of cracking was included in the dataset in terms of a percentage of the
subsection area. A number of cracking types were defined in Chapter Two and, although
structural failure is largely associated with alligator cracking, all information regarding
cracking types (e.g. alligator, transverse, longitudinal, and parabolic cracks) reported in the
State Highway LTPP database were included in the research dataset. Further manipulation of
the total cracking produced a crack rate (per year) for each subsection, as well as the number
of years of continual cracking (i.e. years with no natural closure of the cracks or maintenance
Shear Data:
Shear failure often manifests in the form of shoving on the pavement surface or, secondarily,
by potholes (refer Chapter Two). Therefore, pothole information (number, depth, and
58
Chapter Three: METHODOLOGY
diameter), shoving data, structural patches, and mechanical damage were obtained from the
State Highway LTPP database. To ensure an unbiased comparison, the latter three data
elements were converted into a percentage of the total subsection area. Pothole depth and
diameter remained in millimetres and the number of potholes remained as a number, as this
condition data type is more meaningful in the raw data format, as per the LTPP programme
collection process.
conventional pavement design practices. For rutting, this was defined as a rut
failure.
The objective of the LTPP programme restricts the use of maintenance until such time that
the pavement becomes unsafe and intervention is necessary to ensure the functionality and
safety of the road pavement. As a result of this, this research defined failure in the State
Highway LTPP dataset as the time of maintenance intervention. Maintenance records and
inspection reports provided the information to establish the type of failure, the year of failure,
and the maintenance treatment applied to the pavement. In the State Highway LTPP dataset,
binary classifiers represent the failure condition of the road pavement sections, with zero (0)
indicating a sound pavement (a positive outcome) and one (1) indicating a failure (a negative
59
Chapter Three: METHODOLOGY
outcome). Given the information available, this dataset defined failed sections for each of the
Much like the State Highway network, the LTPP programme includes sites on the New
Zealand Local Authority network. A total of 82 sites in rural and urban areas of the network
have been monitored since 2003. The data available in the database is identical to that of the
State Highway LTPP network, meaning the format of the data is much the same to the State
Highway dataset. However, Local Authorities manage these sites and consequently the
inventory records are less complete and reliable compared to those included in the State
Highway dataset.
This dataset was utilised in the final module of the methodology (testing the prototype
system) because of the similarities in the presentation and format of the data to that of the
independent dataset was simplified. However, some factors included in the prototype system
were absent in this dataset because of the differences in the road structure. For example, the
road pavements in the Local Authority dataset lack a sub-base layer because the traffic
demand is much less than that of the State Highway network. These differences were
considered minor so to not affect the verification and testing of the prototype system.
Based on the descriptions of the LTPP data presented above, Table 3-1 summarises the
variables included in the computational models for each of the three failure modes. The data
60
Chapter Three: METHODOLOGY
available on material properties was considered to be incomplete and therefore was not
In order to quantify the probability of road pavement failure, the development of the
into a computational model, and probability theory was referenced to combine the
individual failure probabilities into one overall assessment of the failure probability. To
61
Chapter Three: METHODOLOGY
achieve the research aim, this chapter has described a systematic approach comprising of five
stages. These sequential modules were required to produce knowledge of road pavement
failure and failure charts, explore and assess a number of modelling techniques in order to
select an appropriate technique, further develop the overall system, and test the prototype
represent road pavement failure, it was necessary to assess the performance of each and the
requirements of the overall system in order to select the most appropriate technique. The
research dataset from the LTPP programme was used in developing the prototype system
(Chapters Four to Six), with the testing of the system using an independent LTPP dataset
(Chapter Seven). Detailed descriptions of each module are presented in Chapters Four to
Eight.
62
Chapter Four
UNDERSTANDING ROAD
PAVEMENT FAILURE
Complementing the Subject
Knowledge
4.1 Introduction
This chapter addresses the first objective of the research by developing a comprehensive
understanding of road pavement failure that will form the foundation of the model
development (as detailed in Chapter Five). The knowledge obtained will improve the success
of the computational model, as it will ensure the model reflects the knowledge of road failure
as opposed to coincidental relationships in the data that may result from statistical processes
(see Section 4.1.1). Furthermore, it establishes the relationship between the factors involved
in each failure mechanism and determines the causes of failure, thereby facilitating the
63
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
identification of the most probable (critical) failure path assuming all failure paths are
can be used to aid in and inform the development of the computation model;
The conceptual design (refer Section 3.1.1) of the proposed system includes a diagnostic
Since this element is fundamental in the development of the prototype system, understanding
the failure mechanisms was seen as a major part of the research. As road failure is
complicated and difficult to artificially replicate with computational models (Reigle, 2000),
such models.
In practice, diagnosing road failure is challenging for asset managers due to the variable
behaviour of road pavements, the complex interactions of factors that can contribute towards
failure, and the occurrence of multiple failure modes. An incorrect failure diagnosis leads to
ill-chosen and ill-timed maintenance decisions, as well as variable outcomes of the forecasted
maintenance programme and budget. However, a better understanding of the causes and the
identification of failure paths, associated with the underlying causes of failure, can facilitate
the application of appropriate maintenance. Therefore, the diagnostic framework not only
64
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
contributes towards the outcome of this research, but also has a direct practical application
mechanism(s);
• Provide information that can aid in making informed decisions regarding the
• Assist with the interpretation of the results from the computational model, and
• Validate the outputs of the statistical (data driven) computational model based on
Developing such an understanding for the three failure mechanisms (refer Section 1.3.3) of
failure;
The knowledge obtained from these sources was presented in the form of failure charts as
described below. Initially, from first principles surrounding pavement design, a generic
65
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
pavement failure chart was developed and, based on five factor groups identified by experts,
additional failure charts were developed for each of the three failure mechanisms.
The development of the failure charts was based on the FTA approach, whereby a tree-like
chart is used to represent the multiple causes of failure and failure paths are used to show the
collection of factors contributing to failure (Yuhua and Datao, 2005). The breakdown of each
failure mechanism in this way enables the model definition to represent concurrently
occurring failure factors, competently and effortlessly (Lindhe et al., 2009; Patev et al. 2005;
Remenyte-Prescott and Andrews, 2008). The FTA method recognises the importance of
Among other applications, FTA is commonly used in research studies as a diagnostic tool
(Volkanovski et al., 2009) to identify the causes of a specified event. Because of its
prioritisation attributes in the design of fault trees, the placement of highly influencing factors
earlier in the tree gives these factors more control on the outcome (Ortmeier and Schellhorn,
2007). The use of combinatorial logic, such as AND and OR functions, enables the
interactions between the factors to be easily represented (Contini and Matuzas, 2011; Lindhe
et al., 2009) and, as a result, the model can represent a number of scenarios (Remenyte-
Prescott and Andrews, 2008; Yuhua and Datao, 2005). The development of such fault trees
66
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
understanding charts for the pavement types (thin, flexible, unbound, granular
From engineering fundamental design principles, pavement failure can occur through either
the failure of the structure to support the design load (bearing capacity) and / or because the
environmental and traffic load exceeds the design loading (loading demand). Figure 4-1
presents the failure chart developed for any pavement structure, based on the FTA approach
excessive traffic loadings on a poorly designed pavement, the FTA approach accounts for
both of these failure situations. Table 4-1 lists five main groups of factors, which are
considered to contribute towards failure, that were obtained through engineering knowledge
and the use of Figure 4-1. This table successfully summarises the numerous independent
The failure chart presented in Figure 4-1 demonstrates the causative nature of the factors
contributing to generic pavement failure. The chart is presented in such a way that the causes
contributing to the failure are sequential; for example, a pavement with insufficient layer
thickness results in a poor pavement composition, which in turn negatively impacts on the
67
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
The loading demand branch in Figure 4-1 consists of two paths; one associated with traffic
and the other with environmental loading. In the case of the traffic branch, solely traffic
exceeding the design loading can cause failure. Alternatively, failure can be caused by a
combination of excessive traffic and poor pavement design. The environmental loading
branch shows that excessive environmental loading can be caused by cyclic temperatures,
Bearing capacity failure is a result of poor design, the overall quality of the construction of
the pavement, climate or environment factors, or issues associated with the properties of the
subgrade. These causes can be further broken down into contributing causes as presented in
Figure 4-1, for example poor pavement design occurs because of inadequate strength, poor
68
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
Table 4-1 describes the contributing role each of the five identified failure factors (from
Figure 4-1) has on pavement failure. Furthermore, surface condition, although primarily
considered a measure of failure and not a contributing factor, may also be in fact associated
with the occurrence of a particular defect that can induce the development of another defect.
For example, severe rutting can also induce pavement surface cracking, yet the primary cause
of failure in this case would be rutting (Henning, 2008; Martin, 2008); or vice versa.
69
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
From hereon, using the concepts of the generic pavement failure chart demonstrated above,
the remainder of this chapter focuses on three predominant failure mechanisms seen on New
chipsealed pavements. Accordingly, in order to develop the associated failure charts, the
• The canvassing of expert opinions to identify the causes of multiple failures and
Rutting Failure:
other words, the pavement is no longer able to dissipate the forces induced by the traffic
adequately. Rutting appears on pavements as depressions in the wheelpaths. There are two
main types of rutting; one related to the surface layer and one related to the structural layers
of the pavement. This research focuses on the latter, where the pavement presents wide
shallow ruts (CSRA, 1992). The impact from traffic loadings causes the lower layers of the
pavement to compact under the wheel loads (Papagiannakis and Masad, 2008), although
rutting can also be attributed to poor pavement support (CSRA, 1992). Several problems with
the underlying pavement layers, drainage, and aggregates contribute to poor pavement
70
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
support and densification of the materials. Therefore, excessive traffic loadings or problems
with the underlying layers of the pavement are the main causes of rutting failure.
Cracking Failure:
There are a number of different types of cracking (refer Section 2.2.2). However, the main
concern structurally for flexible pavements with all types of cracking is that the formation of
cracks permits the ingress of water into the lower layers of the pavement. Further concerns
include the impact on road user costs given the disintegration of the pavement surface,
resulting from interconnected cracking, which brings about an increase in the roughness
Fatigue cracking is the prevalent type of cracking on New Zealand roads, and is caused by
excessive traffic loadings (CSRA, 1992), but it can also be a result of unbalanced pavement
layers (e.g. the stiffness of the individual pavement layers is unbalanced9 across the
pavement), poor pavement support, and / or the use of brittle surface materials (e.g. asphalt
which has aged excessively) (Henning et al., 2006; Martin, 2008). Narrow carriageways or
those without shoulders, common in rural areas of New Zealand, are also prone to edge
cracking, resulting from a lack of edge or shoulder support from the pavement, given the
shoulder is generally designed weaker than the pavement itself (Thom, 2008).
Joint or imbalanced cracking does not often occur on the roads in New Zealand; however,
most of this type of cracking is caused by a stiff upper layer (Henning, 2008), and usually
occurs in stabilised pavements. Other causes include a brittle surface material, thermal cycles,
9
Where there is an excessive accumulation of stress between successive pavement layers.
71
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
Thermal cracking results from temperature changes in the environment but can also be
caused by surface shrinkage, reflective cracks originating beneath the surface, and top down
The other common type of cracking in New Zealand is longitudinal, transverse, or block
cracking. The causes of this type of cracking is similar to other types of cracking, such as top
down cracking, brittle materials, reflective cracking from underlying layers, and shrinkage. In
Shear Failure:
Shear failure commonly manifests itself on the pavement surface as potholes or deformations
and, unlike rutting, can occur anywhere on the pavement section. Generally, as a result of
seasonally changing zones of moisture in the base layers, shear failures are prevalent on or
near the edge of the pavement (Emery, 1992; Schlotjes et al., 2009; Thom, 2008) (see Figure
4-2 below). Like cracking, such defects can allow water to enter the pavement.
A weak basecourse or underlying layers cause shear failures, due to layer compression,
shoulder support, and material shear can also cause shear failures (Schlotjes et al., 2011).
This type of failure is usually not a direct result of excessive traffic on low volume roads
(Schlotjes et al., 2011); however, traffic loadings can further exacerbate it. Insufficient
shoulder support permits shoulder failure and subsequently results in failure near the edge of
the pavement.
72
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
Canvassing the opinions of experts provided this study with additional knowledge of road
pavement failure. Asset managers from New Zealand RCAs and a pavement specialist highly
regarded in the industry were consulted in this stage to assist with defining the interactions
between factors, infrequent and site-specific failure causes, the occurrence of multiple failure
mechanisms, and maintenance treatments. It is common for multiple factors and / or failure
mechanisms to occur simultaneously, thus making the correct failure mechanism(s) difficult
to diagnose (Schlotjes et al., 2011). Consulting with experts allowed such elements of failure
to be better understood and defined, and therefore the knowledge from the experts was
This section summarises the results of the preliminary data analysis (see Appendix B). The
The above relationships, trends, and factors were not immediately identifiable in the
literature. However, the results of the descriptive data analysis (see Appendix B) suggest the
importance of defining such associations. Given the variation in the geography and
environments of New Zealand, the inclusion of site-specific factors and apparent trends on
New Zealand road networks ensured this research produced comprehensive failure charts,
73
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
including the factors influencing pavement failure and their associated interactions for
The data analysis used simple descriptive statistical tools, included in the R statistical
package (Everitt and Hothorn, 2006), to assess the relationships between the variables in the
datasets considered (refer Table 3-1). Smoothing the trend line when examining the
relationships between the variables was required to remove the noise associated with some
variables and provide clarity to the exercise. In these cases, to avoid confusion, the smoothed
trend line is presented in the figures below as a red line, unless otherwise stated.
The SDC is a local authority at the southern base of the South Island, New Zealand. As the
RCA for the Southland district, the SDC manages a network of approximately
6,450 kilometres of pavements with 2,770 kilometres of these pavements being sealed (Land
Transport New Zealand, 2008). These roads are typical of network hierarchical level and
omit any State Highway roads located in the Southland District. The SDC dataset includes
network data typical of all RCAs in New Zealand and, although the data was not as detailed
in comparison to the research-based LTPP datasets, the SDC dataset is one of New Zealand’s
more suitable local authority datasets, with a completeness of inspection data and an
extensive and inclusive history of maintenance and inspections. The available RAMM data
was able to be collated and manipulated accordingly to obtain variables similar to those
74
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
relationships, such as those involving detailed pavement inventory data. Therefore, the LTPP
datasets, described in Section 3.3, were analysed to ensure such detailed relationships were
included in the development of the failure charts. The two LTPP datasets (State Highway and
Strong correlations emerged between traffic loadings and the base layer properties of the
pavement with rutting failure respectively, such as greater rut depths were reported as the
traffic (AADT and HCV) increased and on older pavements. Although the SDC dataset
showed less robust relationships between rutting and traffic than the LTPP dataset (see
Appendix B), traffic loadings and base layer properties were concluded as predominant
However, negative straight-line correlations were found to exist between high speed rutting
condition data and pavement width in the LTPP dataset, particularly concerning the left-hand
wheelpaths. This suggests the width of the pavement influences the rut depths on the
pavement, as narrower pavements concentrate the traffic into distinct wheelpaths resulting in
less traffic wander across the pavement and a greater amount of deterioration in the
wheelpaths. In addition, the State Highway sites showed greater rutting deterioration than the
Local Authority sites, which is understandable, considering the difference in design life,
traffic loadings, and compositions between these two networks. Conclusive relationships
concerning rutting in the right-hand wheelpath were evident in this analysis, in comparison to
the left-hand wheelpath, although it should be noted that the reported rutting depths were
greater in the left-hand wheelpath. Figure 4-2 demonstrates how the seasonal variation in
75
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
moisture directly underneath the left-hand wheelpath impacts on the robustness of the
Furthermore, by visual inspection of the smoothed relationship (refer red line in Figure 4-3),
pavements with cracked surfaces, predominantly alligator cracking, showed greater rut depths
than those without cracked surfaces; however when other cracking types were explored, there
was no distinct differences in rutting depths between these two surface conditions. From this,
alligator cracking can be seen to aggravate the progression of rutting, most probably as a
result of water ingress into the lower pavement layers further deteriorating the pavement’s
76
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
Cracked
Not Cracked
25
9-20
10
5
25
20
15
0-9
10
5
AADT
Figure 4-3: The influence of cracked surfaces on rutting depths, based on alligator cracking only,
for the State Highway LTPP road network
From correlation plots, greater rut depths were visible on older pavements compared to
younger pavements, and a rapid increase (shown by the red trend line associated with the
positive rut progressions in Figure 4-4) in the rut depths were observed as older pavements
aged further. The blue trend line in Figure 4-4 presented little information on the relationship
between negative rut progressions and the age of the pavement. However, on young, weak
pavements that carried increased traffic volumes, the pavement was found to be more
susceptible to rapid rutting deterioration, yet older pavements did not. Interestingly and
unexpectedly, thicker pavements (100 – 1000 mm deep) in the LTPP datasets displayed
greater rut depths than thin pavements (0 – 100 mm thick) (see the smoothed red trend lines
in Figure 4-5), although this may be a direct result of the pavement design. Capturing this
77
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
anomaly in the failure charts poses difficulty; instead, this research included pavement layer
LEGEND
Positive Rut Progressions (mm/yr)
Negative Rut Progressions (mm/yr)
Mean Rut Progression (mm / yr)
2
1
0
-1
20 40 60 80 100
Figure 4-4: Aged pavements in the State Highway LTPP road network showing greater rut progressions
Thick Thick
Thin Thin
Rutting Depth in the Right Wheelpath (mm)
Rutting Depth in the Left Wheelpath (mm)
20
20
15
15
10
10
5
5
AADT AADT
78
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
From the above, it may be seen that the following factors influence rutting failure:
• Pavement width;
• Cracked surfaces;
• Base layer properties, namely the age and thickness of this layer.
The correlation plots from the analysis on the SDC dataset suggested that the cracking
reported on the pavement was predominantly caused by traffic related factors, including
traffic loadings. The SDC dataset provided limited information on the different types of
surface cracking; however, a visually apparent relationship was shown between transverse
cracking and the age of the pavement surface in the LTPP dataset, as shown in Figure 4-6.
The analysis of the LTPP data did not show all of the typical relationships expected between
the occurrence of alligator cracking and the factors that the literature suggested may cause
such defects. For example, no conclusive relationship could be found between the age of the
pavement’s surface layer and alligator (fatigue) cracking. Furthermore, alligator cracking was
observed to be more severe across the whole pavement than that observed to occur
specifically in the wheelpaths, which contradicts the expectation that repetitive traffic
loadings can cause alligator cracking. Therefore, it may be concluded that in low volume
roads the predominant causes of alligator cracking is environmental loading or the properties
79
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
0.5
Fraction of Pavement Cracking is Displayed
0.4
0.3
0.2
0.1
0.0
5 10 15 20
Figure 4-6: Relationship between transverse cracking and surface age on pavements in the Local
Authority LTPP road network
failure and parameters that the literature would suggest are influential, due to a lack of
construction of the road pavement. However, the SDC data confirmed that the narrower
which further substantiated the literature detailing the relationship between shear failures and
80
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
The findings from the preliminary analysis reported above were used in the development of
• Traffic factors (loading volumes and vehicle classes) were main contributors to
• Cracked surfaces permitted water ingress and resulted in greater rut depths;
• The age and thickness of the base layer impacted on the depth and progression of
rutting, and
• Cracking was not directly related with the age of the pavement, yet showed
Other notable findings from the analysis of the datasets are summarised below:
• State Highway pavements in the LTPP data were of a higher strength than the
Local Authority pavements, which was expected given the higher traffic volumes
• The design of the State Highway and Local Authority pavements differed as a
result of the diverse road environments and traffic loadings of these pavements.
Given this, opposing correlations were drawn from the analysis of the two LTPP
81
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
(State Highway and Local Authority) datasets. Although, in these cases, this
respective causes, the research datasets provided valuable knowledge to the failure
The failure charts presented in Figures 4-7 to 4-11 have been developed, following the
principles of typical FTA10 (Lazzaroni et al., 2011) for creating fault trees, for rutting,
cracking, and shear failure using the engineering knowledge ascertained from the above
• The results from the analysis of the LTPP datasets and SDC dataset, where
The presentation of the failure charts follows that of Figure 4-1 (refer Section 4.3) where the
factors involved in pavement failure are sequential, based on their causative nature. Figures
4-8 to 4-10 show the five predominant types of cracking (fatigue, edge, joint or imbalanced,
thermal, longitudinal or transverse or block cracking). Based on the FTA approach (refer
Section 4.2), the failure charts present a combination of factors, indicating that failure is not
always a result of a single factor as shown through the use of multiple and successive
10
Such principles for the development of fault trees include defining the problem, setting rules for the
constituents of the failure on whether the causes of failure can occur simultaneously or distinctly, and event
ordering.
82
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
branches. These combinations represent the failure paths, which in turn identifies the
predominant causes of failure. The individual failure paths consider the interactions between
the failure factors, such that the interdependent nature of the factors involved in pavement
failure can be represented in the computational model (see further Section 4.6).
For example, in Figure 4-7, rutting failure can be solely attributed to excessive traffic
loadings, or attributed to a larger number of factors stemming from poor pavement support
under the induced (design) traffic loadings. However, poor pavement support is a product of a
weak basecourse, arising from water ingress as a result of inadequate drainage for example.
Therefore, the cause of rutting in this case is inadequate drainage which, defined by the
critical failure path, is a result of water ingress weakening the basecourse and ultimately
resulting in rutting.
Examples of rutting not directly caused by traffic loadings are shown under the deformation
branch in Figure 4-7, which can be further split into defects contained in the structural layers
of the pavement. These defects can either be associated with the base layer or the subgrade.
The base layer problems result from inadequate pavement design, quality of the construction
and materials, excessive traffic loadings, or narrow carriageways. The problems with the
subgrade result from inadequate pavement design, the design life of the pavement above the
subgrade has been exceeded, excessive traffic loadings directly affecting the subgrade, or
excessive moisture content in the subgrade to cause deformations to the pavement. These
defects are further defined, as shown in Figure 4-7. Similar approaches to failure for cracking
It is recognised that failure could occur with failure modes occurring simultaneously (for
example rutting could occur at the same time as cracking). Such interdependencies have not
83
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
been represented directly on the failure charts to preserve the clarity of the two-dimensional
failure paths. Nevertheless, the developed computational model is capable of dealing with
84
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
85
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
86
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
87
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
88
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
89
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
Figure 4-12 illustrates conceptually how the three sources of knowledge captured in the form
of failure charts is used in this research to diagnose the cause of pavement failure and
subsequently to assist in determining the probability (likelihood) of failure. Given that the
five failure factors (traffic, composition, strength, environment, and subgrade sensitivity) are
embedded in each failure chart, the combinations of these factor groups are presented as the
failure paths involved in each failure mechanism, per failure chart, and therefore represent the
interactions between the causes of failure for each failure mechanism. Each of these factor
combinations will be modelled (see Chapter Five), and the successful factor combinations
will be compared against the failure charts to ensure the inference of engineering knowledge
in the computational model (refer Section 3.1.1). Only the successful factor combinations
associated with the respective failure mechanism will be included in the development of the
prototype system (see Chapter Six). With this approach, the combinations taken forward in
this research will represent each failure path presented in the developed failure charts,
As the most probable failure path is predicted by the system (refer Section 3.3.4 and further
Section 6.2.1), the knowledge captured in the failure charts is used to further diagnose the
90
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
Data
FAILURE
UNDERSTANDING Data
Modelling
This chapter discussed the importance of, and the process associated with, fully
understanding road pavement failure. To this end, this research developed a methodology of
Five groups of factors were identified in this chapter that influence pavement failure, which
could be considered generic for sealed road pavements. This chart was then expanded and
further developed using the FTA approach to describe rutting, cracking, and shear failures
(Figures 4-7 to 4-11). This process was further informed by the analysis of two datasets
obtained on New Zealand roads, ensuring that site-specific failure knowledge was captured in
Although the focus of these charts is on flexible pavement failure, the methodology
91
Chapter Four: UNDERSTANDING ROAD PAVEMENT FAILURE
Complementing the Subject Knowledge
The failure charts developed, and the engineering knowledge which they capture, will be
used to inform the development of the computational model which will ultimately be used to
determine the probability of pavement failure from road datasets. The choice and
92
Chapter Five
A COMPARATIVE STUDY OF
CLASSIFICATION
TECHNIQUES
Which Technique is Best?
5.1 Introduction
To address Objective Two of the research, this chapter presents a comparative study of a
number of classification techniques suitable for the prototype system together with an
objective process for selecting the most appropriate modelling technique for the prototype
system. The literature review (Chapter Two) identified a number of modelling techniques that
modelling. However, given the definition of failure in the research dataset (refer Section
93
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
problems, and these studies were considered in the development of this chapter.
The purpose of the comparative study is therefore to comparatively assess the performance of
performance elements, are used to select a number of suitable classification techniques for the
task at hand from the methods reviewed in Section 2.6. These techniques are discussed in
Utilising a second criteria, the techniques are ranked based on their performance,
interpretability and usability, and model generalisation to determine the most appropriate
The methodology presented in this chapter was divided into three sections, as shown in
Figure 5-1:
literature review using the first criteria, which may be considered initially suitable
measures, and
3. Rank each technique, using an additional set of criteria, to determine the most
94
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
The first criteria referenced the advantages and disadvantages of each classification technique
reviewed in Section 2.6 (refer Table 2-3). The criteria included the performance of the
technique, user interpretability, short running time, and specific criteria to meet the needs of
95
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
road network asset managers, such as the inference of engineering knowledge into the model,
non-linearity properties, the use of the model with a large number of input variables, and non-
however, it is often characterised as a poor performance measure when used alone, due to its
biasness towards the results (Ben-David, 2008; Parker, 2011). Much of the literature still used
accuracy to evaluate the techniques, given its simple calculations, often alongside other
performance measures. Therefore, this research employed accuracy alongside other point
performance measures, which are further described in Section 5.3.2, and listed below:
• Accuracy;
• Misclassification error;
• F-score, and
• Phi coefficient.
The second criteria ranked each of the suitable techniques according to their performance in
three categories:
• Generalisation of the model to ensure the technique is not overly complex to the
96
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
This criterion above is weighted based on the importance of each category to the
implementation of the overall prototype system in the asset management industry of road
pavements.
R statistical software (Everitt and Hothorn, 2006), previously used in Chapter Four, was
The State Highway LTPP dataset described in Section 3.3.1 was used in this chapter.
Forecasting the future probability states of the LTPP sites was not addressed in this study;
instead, to increase the amount of data available to train the modelling technique with, the
historical data collected in the LTPP programme was used inherently as single entries
resulting in a total of 4512 datapoints in the research dataset. Table 3-1 presented earlier
details the variables from the State Highway LTPP dataset to be included in each of the
models, per factor group identified in Chapter Four (refer Table 4-1). It should be further
noted that the comparative study herein focuses only on the rutting, fatigue cracking, and
For success in modelling and to remove any discrepancies in the research dataset, the State
Highway LTPP dataset was manipulated consistently across all variable fields prior to the
1. Missing Values: Where missing values occurred in the dataset, a nominal value
of zero (0) was assumed. This represents the worst possible scenario for the road
pavement consistently across the majority of the missing fields in the State
97
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Highway LTPP dataset. For this study, this assumption was considered sufficient;
in the dataset differ significantly. For example, the traffic data ranges from
between 100 and 10,000 vehicles per day, while the pavement age variable has
values which are seldom above 100. Therefore, to avoid biasness between the
independent variables and minimise data redundancy within the model, each
Adopting this assumption often carries concerns given road pavement data seldom
preliminary analysis of the State Highway LTPP dataset did not indicate a
therefore, this assumption will not impact negatively on the purpose of this
result, data representing sound road sections in road network datasets occur far
more commonly than data representing failed sections, resulting in limited failure
data with which to train a computational model. To address the issue of imbalance
between reported failures and sound pavement sites, this research applied a
98
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
the State Highway LTPP dataset. This is further discussed in Chapter Eight (see
Section 8.4.2.3).
For the comparative study, suitable classification techniques were selected using the first
criteria, based on the merits and limitations of both discriminative and generative techniques
(refer Table 2-3). Section 2.6 reviewed the literature where these techniques have been used
as a binary classifier. The first criteria inferred non-technical aspects of the methods with the
1. The overall prototype system must accurately classify the occurrence of failure for
the choice of the most suitable technique for further development and, therefore,
binary studies must be well documented in the literature despite the research
domain;
result, an extensive amount of data is associated with road networks. The size of
the dataset is greatly influenced by the size of the road network; thus the
modelling technique must be able to cope with all sizes of datasets with an
3. Chapter Four discussed the complexity of road pavement failure, particularly the
complex interactions between failure factors (Reigle, 2000) and the difficulty of
99
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
model function will accurately replicate such behaviour, thus pavement failure
techniques should be favoured in this study; however, if required, the use of linear
RCAs. The time required for the modelling technique to output predictions should
5. Furthermore, the training process of the model should be simple, as well as easily
inferred and based on the occurrences in the data (e.g. no randomness in the
approaches, and stochastic processes in the model, as well as being simple and
easily inferred;
6. The overall process of this model, from training and application, should be easily
7. The novelty of this research infers engineering knowledge into the model
development and, therefore, the modelling technique should be able to make use
Table 5-1 evaluates each of the techniques identified in the literature review, based on the
100
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Criteria
Learning phase
Interpretability
‘YES’ TOTAL
Extensive data
Short running
Human Input
Non-linearity
Performance
of the model
of the model
and inputs
time
Modelling Approaches
Bayesian networks NO YES YES NO NO1 YES YES2 4
Decision (probability) trees YES YES YES YES YES YES YES 7
3 4
Hidden Markov model NO YES YES YES NO YES NO 4
5
k-nearest neighbours NO NO YES NO NO YES YES 3
Linear discriminant analysis NO NO NO YES YES YES NO 3
Logistic regression YES NO NO YES YES YES NO 4
6 7 8 9
Naïve Bayes NO YES YES YES NO YES NO 4
10
Neural Networks YES YES YES NO NO NO YES 4
Random Forests YES YES YES NO YES NO YES 5
11
Support Vector Machines YES YES YES NO YES NO YES 5
1 7
Probabilistic or stochastic processes involved He et al. (2012)
2 8
Subjective approach Probabilistic or stochastic processes involved
3 9
Requires a large amount of training data Strong independence in the model
4 10
Unfounded assumptions included in the model Supervised learning
5 11
User defines the distances between neighbours Through the use of kernels
6
Inferred from Bayesian networks
From the results above, the following five techniques with the highest score were selected to
further evaluate:
• Logistic regression, out of all the techniques with a score = 4, the performance
• Neural networks, out of all the techniques with a score = 4, the performance
criteria dominated.
101
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
al., 1997). However, the nature of road pavement data is not always linearly separable;
therefore, to address this, non-linear techniques were considered. Support vector machines
met five out of seven criteria above and neural networks was reported a good performer in the
literature. Both methods infer human knowledge into the model and are non-linear; however,
neural networks are less interpretable than support vector machines, although the way in
which both of these techniques work can be difficult to understand (refer Table 2-3).
Decision (probability) trees outperformed the other classifiers across all the criteria, yet the
simple format of this modelling technique suggests a limited performance, through a simple
model curve (decision boundary), associated with the optimal solution. However, random
forests are more expressive but very difficult to interpret visually in comparison to trees.
As previous comparative studies have focussed on the performance of classifiers given data
outside of the transportation sector, the five selected techniques are further scrutinised using
and
102
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
The previous section selected five techniques for further investigation from those identified in
the literature review. This section reviews each of these five techniques with respect to the
Logistic regression is a linear classifier and notably reported as one of the simplest classifiers,
Tu, 1996). Linear relationships between the independent and dependent variables are easily
explained with this method; however, given the general non-linearity of road pavement data
(refer Chapter Four), transforming the data by a natural logarithm of the odds, or logit of the
dependent variable (Equation 5-1), can illustrate non-linear relationships (Everitt and
8
Equation 5-1
6
7 = ln
// = ln = 9 + :ଵ ;ଵ + ⋯ + : ;
1−8
The outputs of the logit(Y) would be of a natural logarithm and therefore it is possible to
convert the predicted outputs into a feasible solution, the odds, by the use of exponentials
Equation 5-2
//(7) = ୪୬ଵିగ =
గ
ఈାఉ ା⋯ାఉ
11
Where the variable can be divided into two parts or classes. Binary variables are characterised by dichotomy.
103
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Converting the output for odds (Equation 5-2) into practical probabilities, Equation 5-3 is
Equation 5-3
//(7)
7 =
7 = 1 =
ఈାఉ ା⋯ାఉ
Equations 5-1, 5-2, and 5-3 express the exact same logistic regression output in three
different formats, depending on what is desirable by the user (Menard, 2010). The easiest
method to analyse a binary dependent variable, given the non-linearity, is using the logit form
of the probability. Mathematically, the logit format allows the output to be ± ∞ which, given
However, the transformation given in Equation 5-3 ensures the predicted probabilities adhere
The optimal model curve (the computational solution to the mathematical problem) is found
by an iterative process, which revises the initial provisional solution until the errors of the
model are no longer improving or are negligible. The parameters associated with the
converged (best) solution characterise the trained logistic regression model (Menard, 2010).
Neural networks are a non-linear classifier that have performed well in solving classification
and prediction problems, yet they notably locate local minima as opposed to the global
minimum (Eastaugh et al., 1997; Faraway, 2006). Unfortunately, neural networks can be
difficult to interpret and are considered a ‘black box’, particularly as the connections from the
inputs to the outputs through the hidden layer(s) are unknown to the user (Tu, 1996). Despite
the limitations of this method, it is a valuable computational tool and is popular for solving
104
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
2. The learning process of the network for updating the weights, and
3. The activation function for converting the weighted input into the predicted
output.
minimum of three layers; input(s), output, and at least one hidden layer. Each node in one
layer is connected to every node in the succeeding layer by a weight, as shown in Figure 5-2.
With similarities to the human brain, the inputs are processed through a series of nodes
arranged through hidden layers, organising themselves unbeknown to the user in such a way
that the desired output is predicted. Although the presence of a hidden layer infers a ‘black
box’ approach, the inclusion of this layer allows a neural network to model non-linear
relationships (Tu, 1996). Multiple hidden layers are possible in neural networks; however,
overly complicated and large networks may result in over-fitting of the data, such that there is
no advantage of having more than one hidden layer (Tu, 1996), or more than 10 nodes in one
hidden layer.
105
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
The training of this network is done by the backpropagation learning process, popular with
neural network applications to converge the best solution. Initially, the network error is fed
backwards through the network (from output to input) to adjust the weights of each
connection line between the nodes (Aitkin and Foxall, 2003; Tu, 1996).The network error
Equation 5-4
= * ଶ
1
2
This error is then sent back through the network and each hidden node (towards the input
node i) to adjust the connection weights < (Tu, 1996), using a gradient descent (Equation 5-
5), where = defines the learning rate (always positive) to ensure the convergence of the
results (Gupta and Lam, 1998; Reed and Marks II, 1999).
>
Equation 5-5
><
Δ< = −=
,- ,-
ℎ: − = /′(. ) 0 − 1
,. ,.
The desired output of a node is defined by the activation function (?). A linear activation
function would remove the hidden layer of the network resulting in a linear system, thus such
A number of functions are possible for neural networks; however, for this type of multilayer
perceptron network the sigmoid activation function is appropriate (Equation 5-6) and can be
106
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
? = tanh
@ / ? =
1 + ିజ ିଵ
Equation 5-6
The normalisation and differentiable rules associated around Equation 5-5 are directly
transferable to the activation function; the sigmoid function satisfies these rules and is
therefore applied to all the nodes (Aitkin and Foxwall, 2003). After the error (Equation 5-4)
converges, the resulting neural network model is defined by the connection weights (shown
in Figure 5-2).
A large number of studies have employed support vector machines as a non-linear classifier
(Rogers and Girolami, 2012) since they were first developed by Cortes and Vapnik (1995)
and may be considered to be similar to neural networks. Unlike logistic regression, the
outputs are non-probabilistic; however, a distribution can be fitted around the outputs to
obtain probabilities by employing scaling methods (Karatzoglou et al., 2006). Support vector
machines create a non-linear decision boundary to separate the data with minimal errors.
To do so, the method uses a higher dimensional feature space to map the input vector, defined
by the input variables, and it is hoped that in this feature space the transformed data can be
linearly separated, such as in Figure 5-3. To transform the input data into the higher
dimensional space, kernels are employed. Table 5-2 presents the kernels available and their
associated decision boundary shape. A linear kernel will produce a linear separator and a
polynomial kernel will result in a curved decision boundary; thus confirming the polynomial,
radial basis function, and sigmoid kernel are more flexible than the linear kernel. Each of
these requires additional input parameters that are set by the user, thus inferring human
12
Analogous to the human brain, where the synapses are the junctions and connections between the nodes.
107
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
knowledge into the computational model. The radial basis function is a popular kernel to use
because the decision boundary associated with this kernel is easily manipulated through the
transformed data, opposed to the others. When used, although not popular with support vector
machines, the sigmoid kernel replicates a multilayer perceptron neural network (Rogers and
Girolami, 2012).
Figure 5-3: Transformed data using the support vector machines classifier
(adapted from Van Looy et al., 2007)
Description/Decision
Kernel Method Equation
Boundary
Linear Straight line, through the data 23 , 3 = 3
3
Polynomial (of degree d) Curve, through the data 23 , 3 = (1 + 3
3 )
Radial Basis Function (or Circular, surrounding part of
23 , 3 =
( )
Gaussian) the data
Sigmoid S-curve, through the data 23 , 3 = tanh(4 3
3 +
)
The decision boundary is used to classify the data and, as a linear separator, can be defined as
(Kecman, 2005):
# = ±
A # +
Equation 5-7
்
108
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
The ± sign references the data either side of the decision boundary and A and are
calculated from the data in the learning task of the model. The parameters relate to the
optimal model where the margin (γ) defined in Figure 5-4 is greatest, or the loss is
minimised.
The decision boundary is optimally placed to maximise the margin between the closest points
either side to the boundary. This margin is calculated as the perpendicular distance from the
decision boundary to the closest points on either side (Gil and Johnsson, 2011). A smaller
margin would allow for a greater error (loss) of the model with a smaller confidence between
class separators. Therefore, with reference to Figure 5-4, the margin can be written
Equation 5-8
B= A்
Dଵ − Dଶ
1
2‖A‖
109
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Girolami, 2012):
Equation 5-9
A்
Dଵ − Dଶ
1
‖A‖
2B =
A் Dଵ − A் Dଶ
1
‖A‖
=
(A் Dଵ + ) − (A் Dଶ + )
1
‖A‖
=
1 + 1
1
‖A‖
=
B=
1
‖A‖
Therefore, to solve the classification problem, Equation 5-9 must be maximised. Since D and
D are the closest points to the decision boundary, these are known as the support vectors and
define the decision boundary; thus they can be seen as the model definition. To predict a new
observation, instead of the model using the entire dataset to relate the new observation in the
feature space as some techniques do, support vector machines use only the known support
vectors to define the boundary (Rogers and Girolami, 2012). The new observation is plotted
with just this information, making this technique more computer efficient and faster than
others.
Decision trees, referred to in the remainder of the thesis as probability trees, are a hierarchical
classification method to solve both classification and regression problems. This technique is
very interpretable and easily understood, yet their performance is expected to be inferior to
the previous techniques, given the limited expressiveness of the model format and associated
110
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
decision boundary in comparison to the other methods. However, given the advantage of
A sequence of questions is answered at each ‘node’ and when there are no longer any
questions to ask, the tree ends with a terminal node or ‘leaf’. Foote (1994) described the three
1. Splitting rule, which determines the decision threshold of the node (in other
2. Stopping rule, to determine when the recursion ends (when the tree no longer
The data is partitioned between each node by answering each nodular question; however, the
possible positions of these partitions are endless. An optimal partition may not be attainable
and instead the position of the partition is constrained by the use of hyperplanes. The learning
phase of the model involves establishing this data separation and, subsequently, the optimal
position of the hyperplanes. Like most other methods, the splits are chosen to minimise the
loss of the model, or in other words maximise the gain (Foote, 1994), defined by Chen et al.
.
# , states given the inferred information from the model, such that:
+
# , = .
− .(# , )
Equation 5-10
The tree continues to grow until no further split is possible or the node contains only one
class. To avoid over-fitting, some branches with low gain or contribution to the overall model
111
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
can be pruned (Everitt and Hothorn, 2006). This may mean that some internal nodes are
converted to terminal nodes, ignoring any subsequent splits beyond that node. As a result, an
On completion of the learning phase, nodes are assigned a respective class label and the
statistics associated with each terminal node, such as probabilities, are computed (Everitt and
Hothorn, 2006). For a new prediction, the constructed tree assigns the new observation with
the statistics associated with the predicted terminal node, once the observation has passed
Random forests are an ensemble13 technique, analogous to bagging14 except this technique
compounds only the same type of classification model, where a collection of decision
(probability) trees are constructed to create a forest. Since they are constructed from
probability trees, the concepts of trees and recursive partitioning are applicable to random
forests to the extent that the learning mechanism of random forests is that of probability trees.
Although random forests are computationally effective given their substantial expressiveness
becoming less interpretable than probability trees and difficult for the user to understand.
A number of specific rules are followed to construct a forest in addition to the learning phase
of decision (probability) trees. In the learning process, some randomness is introduced into
the splitting rule for each individual tree by growing each tree on a random subsample of the
data (Robnik-Šikonja, 2004). The collation of the trees aims for minimal repetition of the data
13
Several computational models are used together, often one after another, to improve on the overall
performance of the individual techniques.
14
Equal weights are given to each model, in this case each tree developed, to create an overall compounded
modelling system of individual trees.
112
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
classes and information at each terminal node. To achieve this, weighted voting is used to
combine the outputs of the individual trees, where new predictions are determined by the
highest number of class votes from all trees in the forest (Sirikulviriya and Sinthupinyo,
2011).
The definition of the performance of a model is multi-dimensional, such that the assessment
methods of some performance measures are nonchalant and obscured. Logically, accuracy or
measure alone as the bias results do not allow for random successes (Ben-David, 2008;
Dreiseitl and Ohno-Machado, 2002; Parker, 2011). To ensure this comparative study
• Point Measures: These measures estimate the performance of the model at a specific
threshold and ignore all other threshold points. These types of estimates assume that
false positives and false negatives are equally undesirable (Parker, 2011). Examples
changes, such that the consequences with all incorrect predictions are considered in
curve and Cohen’s kappa curve, are calculated by plotting such constituents (e.g.
sensitivity, false positive rate, and Cohen’s kappa) against each other across various
thresholds. Integrating the curve with the associated cost function evaluates the
113
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
performance of the technique using such measures as area under the receiver
operating characteristic curve, the area under the Cohen’s kappa curve, and the H-
With the changes in output demand of classification techniques, integrated measures are now
favoured for assessing the performance of modelling techniques as opposed to point measures
(Parker, 2011). Ideally integrated measures would be used for this research; however,
because of limitations in the data, the costs (consequences) associated with false positives and
false negatives were assumed to be equal, such that it is equally undesirable (e.g. an equal
cost function) to have an incorrect prediction of failure, whether that be a false positive or
false negative prediction. However, Parker (2011) reported that the phi coefficient was a
the F-score. Therefore, the F-score and phi coefficient, in conjunction with accuracy and
misclassification error, were used to assess the classification techniques using a failure
threshold of 0.5 (
; ≥ 0.5 = 1). These techniques are described further below.
Figure 5-5 defines the elements of the confusion matrix, each used in the subsequent
performance measures.
DATA FAILURES
0 1
PREDICTIONS
N1
TP FP
0 True Positives False Positives
Number of predicted
non-failures
N2
FN TN
1 False Negatives True Negatives
Number of predicted
failures
N3 N4 NTotal
Number of non-failed Number of failed Total number of sites
sites sites (or predictions)
Figure 5-5: Confusion matrix (Fawcett, 2006; Sokolova and Lapalme, 2009)
114
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
For this research, sound pavements were assigned a zero (0) class, therefore correctly
predicted sound pavements were represented by true positives and sites predicted to fail but
were in fact sound were defined as false negatives. Failed pavements were assigned a one (1)
class and, therefore, true negatives were correctly predicted failures and sites predicted to be
correctly predicted (what proportion of the predicted sound pavements were in fact
the predicted sound pavements (Fawcett, 2006; Powers, 2011; Sokolova and Lapalme,
2009).
-
Equation 5-11
! =
- +
ℎ: ℎ ! * ( 5 − 5)
Recall (sensitivity, true positive rate): The proportion of sound road pavement
sections correctly predicted (how many sound pavements were predicted as sound?).
The result indicates the ability of the technique to correctly identify sound pavements
(Fawcett, 2006; Powers, 2011; Rogers and Girolami, 2012; Sokolova and Lapalme,
2009).
-
Equation 5-12
! =
- + ே
ℎ: ℎ ! * ( 5 − 5)
115
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Negative predictive value: Opposite to the positive predictive value, the proportion
of the predicted failed sites that are correctly predicted (what proportion of the
-ே
Equation 5-13
ℎ: ℎ ! * ( 5 − 5)
Specificity (true negative rate): The proportion of correctly predicted failures given
the total number of failed sections in the data (how many failed pavements were
technique to identify failed pavements (Fawcett, 2006; Powers, 2011; Rogers and
-ே
Equation 5-14
)E!! =
+ -ே
ℎ: ℎ ! * ( 5 − 5)
The performance assessment of the techniques should not be solely based on the direct
relationships from the confusion matrix as the above tests rarely conclude the practicality and
Both the accuracy and misclassification error measures calculate the percentages of
incorrectly predicted pavement sections, as shown below in Equations 5-15 and 5-16
(Powers, 2011; Rogers and Girolami, 2012; Sokolova and Lapalme, 2009). Ideally, a higher
116
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
percentages of correctly and incorrectly predicted sound and failed pavement sites. However,
as stated above, the results from these calculations should not be used alone in assessing the
∑
- + -ே
Equation 5-15
%!! ! =
்௧
× 100 %
∑|ௗ௧ௗ − ௧௨ | ∑
+ ே
Equation 5-16
5!! =
்௧ ்௧
× 100 % = × 100 %
,
= ,
5.3.2.3 F-Score
The F-score measures the performance of the binary classifier on a scale of zero (0) to one
(1), where the closer the value is to 1, the more accurate the method is regarded (Parker,
2011; Sokolova and Lapalme, 2009). It is a weighted average of the precision and recall
values (the predictive power of the model to sound pavements); however, it neglects the
number of correctly predicted failures (e.g. the true negative rate). It is a preferred measure of
performance over accuracy calculations given its inclusion of incorrect predictions, and is
2×E×
Equation 5-17
_)! =
E+
ℎ:
=
=
117
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
The phi coefficient, also known as Matthews Correlation Coefficient, measures the
agreement between the inputs and the output and, therefore, how well the selected technique
predicted failed and sound pavements. A negative agreement, equal to -1, suggests the
majority of the results are incorrectly predicted, and a positive agreement of +1 would
demonstrate that the method is accurate in predicting road pavement failure (Powers, 2011).
A value of zero (0) indicates no relationship between the predictions and input variables.
The phi coefficient is often favoured above the F-score (Parker, 2011) because it takes into
account negative values (or failed pavement predictions), unlike the F-score which fails to
use the predictions of true negatives (failed pavements). It is calculated using the following
- × -ே − × ே
Equation 5-18
Hଵ × ଶ × ଷ × ସ
∅=
where: T , T = Number of true positive and true negative predictions (refer Figure 5 − 5)
ℎ ℎ ℎ ( 5 − 5)
Using the R functions described in Appendix C, each classification model type was
constructed and tested using a reserved portion of the State Highway LTPP dataset. Given the
size of the research dataset, only a limited number of datapoints were able to be reserved
from this dataset for testing purposes; therefore, to ensure variability of the predictions is
accounted for in the results, a 10-fold cross-validation sampling method was employed
(Rogers and Girolami, 2012). This approach divided the dataset into 10 subsamples of equal
118
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
size, randomly selected, reserving one subsample for testing at all times. Therefore, over the
10 repetitions, 90 % of the dataset was used in the training phase of the model and the
remaining 10 % was reserved for testing. An average of these 10 repetitions, including each
calculated performance measure, was used to evaluate the performance of the classifiers.
With the randomness of this method, the testing sample does not guarantee to include an
Each of the factor combinations, comprising of the factor groups discussed in Chapter Four
(refer Section 4.6), shown in Table 5-3, were modelled individually to:
• Determine the correlation between the factor groups and failure mechanism to
To do so, the accuracy of each model constructed was reported. An investigation of the
distributions of the four performance measures used box and whisker plots (Ayyub and
McCuen, 2003) to establish the overall performance of the technique. Density distribution
plots explored the performance of each factor combination per failure mechanism, against
each technique, to further add to the accuracy results. Finally, the Wilcoxon Signed Rank
Test (R Core Team, 2011) analysed the statistical significance between the means of the
performance measures. For this comparative study, the performance evaluation remained
119
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Table 5-3: Factor Combinations of the Factor Groups
Combination
Environment
Environment
Composition
Composition
Sensitivity
Sensitivity
Factor
Factor
Condition
Condition
Subgrade
Subgrade
Strength
Strength
Surface
Surface
Traffic
Traffic
1 x 33 x x x
2 x 34 x x x
3 x 35 x x x
4 x 36 x x x
5 x 37 x x x
6 x 38 x x x
7 x x 39 x x x
8 x x 40 x x x
9 x x 41 x x x
10 x x 42 x x x x
11 x x 43 x x x x
12 x x 44 x x x x
13 x x 45 x x x x
14 x x 46 x x x x
15 x x 47 x x x x
16 x x 48 x x x x
17 x x 49 x x x x
18 x x 50 x x x x
19 x x 51 x x x x
20 x x 52 x x x x
21 x x 53 x x x x
22 x x x 54 x x x x
23 x x x 55 x x x x
24 x x x 56 x x x x
25 x x x 57 x x x x x
26 x x x 58 x x x x x
27 x x x 59 x x x x x
28 x x x 60 x x x x x
29 x x x 61 x x x x x
30 x x x 62 x x x x x
31 x x x 63 x x x x x x
32 x x x
120
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Based on the predicted probability of failure, the assessment of each classification technique
used the performance measures described in Section 5.3.2 to assist in determining the
Tables 5-4, 5-5, and 5-6 present the accuracy and misclassification error percentages based
on Equations 5-15 and 5-16, for each failure mechanism, and report the successful factor
combinations that failure can be correctly predicted from, given a successful accuracy
threshold of 100 % (and 90 % in the case of shear failure). In addition, the factor
combinations that performed the poorest for each technique are identified. The accuracy and
misclassification error percentages for all factor combinations trialled can be found in
Appendix D.
It is difficult from these results to determine the most accurate modelling technique; instead,
the technique(s) with the smallest misclassification error across the three failure mechanisms
include:
• Fatigue Cracking: Support vector machines and random forests (Table 5-5), and
Table 5-6 shows the performance of all modelling techniques are considerably less accurate
in predicting shear failures than that of the other two failure mechanisms. In addition to the
combinations were considered ‘worst trials’ for the shear failure mechanism, compared to the
121
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
rutting and fatigue cracking analysis. The inherent behaviour of shear failures and the lack of
material information in the State Highway LTPP dataset could be reason for this observation.
The successful factor combinations identified by the support vector machines and logistic
regression techniques correlated well with the respective failure charts in Chapter Four, such
that combinations containing strength, traffic and composition factors, for example, were
associated with predicting rutting failure. However, with the large number of successful
factor combinations reported for neural networks and random forests, the correlation may be
considered coincidental in some cases. Despite this, the surface condition factor group was
only evident in successful fatigue cracking combinations, which suggests forecasting failure
122
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Table 5-4: Summary of the Accuracy and Misclassification Error Results for Rutting Failure
(Average of 10-Fold Cross-Validation)
123
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Table 5-5: Summary of the Accuracy and Misclassification Error Results for Fatigue Cracking Failure
(Average of 10-Fold Cross-Validation)
124
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Table 5-6: Summary of the Accuracy and Misclassification Error Results for Shear Failure
(Average of 10-Fold Cross-Validation)
Strong conclusions on the performance of each technique cannot be inferred from the results;
however, from the above tables, the successful factor combinations can be further used in the
125
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Figure 5-6 compares each technique based on the distributions of the four performance
measures. The low phi coefficient indicated a weak relationship between the input variables
and predicted outputs for shear failure. As this was evident across all five techniques, it was
suggested that the available dataset does not contain enough information, such as pavement
For the other two failure types, Figure 5-6 concluded that neural networks and random forests
perform better than the other three classifiers across all four measures. Probability trees
performed marginally better than support vector machines given the accuracy,
misclassification error and phi coefficient assessments, but ignoring the shear failure
mechanisms the two techniques exhibited equal performance in the F-score. The logistic
regression performance was least successful across all four performance measures, with the
range present for each performance measure and failure mechanism exceeding that of the
other four techniques. The raw data used to construct Figure 5-6 can be found in Appendix D.
126
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
127
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
From this, the classification techniques are ranked in the following order, given the
2. Probability Trees
4. Logistic Regression
Figures 5-7 and 5-8 present the density distribution plot per factor combination for rutting
and fatigue cracking failure mechanisms to compare the performance of each classification
technique, given the input variables. Largely, the results reporting on the comparative
performance of each technique were inconclusive. The full collection of density plots are
Rutting:
In general, the density distribution plots for the rutting failure mechanism show little
difference in the distributions between each of the classification techniques over the four
performance measures. However, the plots focussed on the accuracy and misclassification
error percentages showed that overall the performance distribution of the neural network
technique differs from the other four classifiers, across the majority of the factor
combinations. The plots detailing the F-score and phi coefficient have few factor
combinations where the distributions of each modelling technique differ. Generally, the
accuracy, and the associated misclassification error, density distribution plots for neural
networks differ to the other four modelling techniques (see Figure 5-7), suggesting the
128
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
performance of neural networks is significantly different to the others for predicting rutting
failure.
Figure 5-7: Density plots for the accuracy performance measure for the rutting failure mechanism
129
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Fatigue Cracking:
The density distribution plots exploring the misclassification error for fatigue cracking shows
a distinct difference between logistic regression and the other four techniques. The centre of
the distribution curve for logistic regression tends to be greater than that of the other four
techniques (see Figure 5-8), suggesting a higher misclassification error associated with
logistic regression. Only approximately 10 factor combinations (out of 95 including all a and
b trials) showed some variance between the modelling techniques for each performance
measure.
Figure 5-8: Example of density plots for the misclassification error performance measure for fatigue
cracking failure
Shear:
The results of the density plots further supported the conclusions from Section 5.3.4.2, in
regards to shear failure, where the weak relationships presented in the box and whisker plots
were accounted for. Although many of the plots showed a difference in the distributions,
Section 5.3.4.2 concluded poor correlations between the input variables and predicted output;
therefore, the variation in the distribution plots may be a result of this poor relationship and
130
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
To determine the statistical difference, if any, in the reported performance of the five
difference between the distributions of each performance measure. Specifically, the Wilcoxon
Signed Rank test (R Core Team, 2011) is a non-parametric statistical hypothesis test used on
3. The data is measured on an interval scale but does not need to be normal.
Therefore, a paired two-sided test compared the population means of each of the four
performance measures with a desired confidence interval for this hypothesis test of 95 %.
The null hypothesis is accepted if the p-value from the Wilcoxon Signed Rank test is greater
than 0.05 (i.e. = 1 – 0.95), otherwise it is rejected and the alternative hypothesis is accepted
Tables 5-7 and 5-8 show there is no statistical difference between the performance of support
vector machines and probability trees as H0 is accepted across all four of the performance
measures for the rutting and fatigue cracking failure mechanisms. There is also no statistical
difference in the phi coefficient results between the random forests and neural networks,
131
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
For the shear failure mechanism, there is a difference between each classification technique
for accuracy, misclassification error and F-score measures; however, there is no statistical
difference between the performances of the techniques based on the phi coefficient.
Table 5-7: Results from the Wilcoxon Signed Rank Test for the Accuracy and Misclassification Error
Performance Measures, showing the Acceptance of the Null Hypothesis (H0)
Modelling Techniques
Support
Logistic Neural Probability Random
Vector
Regression Networks Trees Forests
Machines
R C S R C S R C S R C S R C S
Logistic
Regression
Neural
Networks
Accuracy
Support Vector
H0 H0
Machines
Probability
H0 H0
Trees
Random
Forests
Misclassification Error
R – Rutting; C – Fatigue Cracking; S – Shear
132
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
Table 5-8: Results from the Wilcoxon Signed Rank Test for the F-Score and Phi Coefficient Performance
Measures, showing the Acceptance of the Null Hypothesis (H0)
Modelling Techniques
Support
Logistic Neural Probability Random
Vector
Regression Networks Trees Forests
Machines
R C S R C S R C S R C S R C S
Logistic
Regression
Neural
H0
Networks
F-Score
Support Vector
H0 H0 H0 H0
Machines
Probability
H0 H0 H0 H0 H0
Trees
Random
H0 H0 H0 H0 H0 H0
Forests
Phi Coefficient
R – Rutting; C – Fatigue Cracking; S – Shear
The Wilcoxon Signed Rank test results conclude that, for both the rutting and fatigue
cracking failure mechanism, there is no difference between support vector machines and
probability trees. In addition, there is no difference between the distributions of the phi
coefficients for neural networks and random forests. Based on this, the techniques can be
3. Logistic Regression
133
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
The previous section assessed the performance of each of the five classification techniques.
This section further builds on those results and bases the choice of the most appropriate
2. The technique must be quick to run, simple to use, and the overall process should
ensure the performance of the model is not compromised when transferring the
prototype system to other road datasets, and ensure the model does not require
large amounts of data for new predictions (e.g. avoid data hungry models).
Based on the criteria above, the most appropriate classification technique for this research is
not necessarily the one that outperforms the others; instead, the needs of the road
management industry have also been accounted for in the criteria for the successful
5.4.2 Discussion
Table 5-9 lists the five techniques accordingly to the criteria above (refer Section 5.4.1),
based on the results from the literature search (refer Table 2-3) and the performance results
134
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
from Section 5.3.4. Logistic regression satisfied four of the five categories, which was
expected given its linearity and simplicity, suggesting its appropriateness to the problem
performance against the other techniques. To avoid selecting a technique that satisfied the
criterion to the overall system aim. Criterion #1 was assigned the largest weighting of 5, to
ensure a poor performing technique was not chosen. The remaining two criterions were
assigned smaller weightings to reflect their importance of the usability and interpretability of
the system. Criterion #2 was sub-divided so the individual elements of the criteria (run time,
Neural
Logistic Logistic Logistic Logistic
1 Networks and
Regression Regression Regression Regression
Random Forests
Probability Probability
Trees and Trees and Support Vector Support Vector
2 Probability Trees
Support Vector Support Vector Machines Machines
Machines Machines
Random Forests
Logistic Random Probability Probability
3 and Support
Regression Forests Trees Trees
Vector Machines
Random
Neural Forests and Random
4 Neural Networks
Networks Neural Forests
Networks
Neural
5
Networks
135
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
From Table 5-10 below, it may be concluded that the most appropriate modelling techniques
were both support vector machines and probability trees. This finding is analogous to the
accuracy results presented in Section 5.3.4.1. The techniques disregarded include logistic
regression due to its poor performance attributed to the linearity of the model function, neural
networks because of the associated long running times and lack of user interpretability
despite its performance, and random forests due to their somewhat complicated model
Criteria
#1 #2 #3
Interpretability
Performance
Ease of Use
Properties
Over-fit
Speed
Overall
Weightings 5 3 1 1 2 Total Rank
Logistic 25+3+1
5 1 1 1 1 3
Regression +1+2= 32
Neural 5+15+5
1 5 4 5 5 5
Networks +4+10= 39
Support
15+6+3+3+
Vector 3 2 3 3 2 1
4= 31
Machines
Probability 15+6+2+2+
3 2 2 2 3 1
Trees 6= 31
Random 5+12+3+4+
1 4 4 3 4 3
Forests 8= 32
Although probability trees could be easily constructed to follow the developed failure charts
(refer Chapter Four), this research selected support vector machines to develop further.
136
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
technique to further develop in the prototype system. Within this methodology, criteria were
developed to select five classification techniques from the various approaches identified in
the literature review (Chapter Two). A comparative evaluation of the performance of each
technique, used in conjunction with objective criteria, selected support vector machines as the
A framework consisting of seven criterions selected five techniques to trial to ensure the
suitability of each technique to the research. The elements of the conceptual design (inferring
engineering knowledge and predicting the probability of failure) were evident in the criteria
and model construction. The selected modelling techniques to trial were logistic regression,
neural networks, support vector machines, probability trees and random forests.
Four performance measures assessed the suitability of each of these techniques further,
sites (misclassification error), F-score and phi coefficient. As accuracy is a poor measure of a
model’s performance, other measures were explored to ensure a thorough and appropriate
evaluation of the techniques, through the use of box and whisker plots, density distributions,
and hypothesis testing. From this analysis, the performance of neural networks and random
forests exceeded that of the other techniques and the accuracy results identified the successful
factor combinations for each failure mechanism, which correlated well with the developed
failure charts.
The second set of criteria in addition to the performance of the modelling technique further
included an assessment of the running speed, interpretability and ease of use of the model,
137
Chapter Five: A COMPARATIVE STUDY OF CLASSIFICATION TECHNIQUES
Which Technique is Best?
transferability of the trained model to other road network datasets. From this analysis, support
Despite the simplicity and interpretability of probability trees, this chapter selected the
support vector machines approach for the prototype system. In particular, the comparative
study concluded this classification technique to be the most suitable and successful for
modelling the road pavement data included in the State Highway LTPP dataset. The
development of the system using the chosen modelling techniques is the focus of the
subsequent chapter.
138
Chapter Six
DEVELOPMENT OF THE
PROTOTYPE SYSTEM
Predicting the Probability of
Failure
6.1 Introduction
The aim of this research is to produce a diagnostic tool capable of quantifying the probability
(likelihood) of road pavement failure. To this end, Chapter Three presented the conceptual
design of the system, and Chapters Four and Five described the development of the
components of the system, as summarised in Figure 6-1. This chapter therefore presents the
development of a fully working prototype system to address Objective Three, based on the
139
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
conceptual design, incorporating the outputs and models developed in Chapters Four and
Five.
1. Failure charts to diagnose the cause of failure (refer Chapter Four), and
Chapter Five).
Failure Charts:
Failure charts (developed in Chapter Four) represent the diagnostic element of the model
design. These charts identify the possible failure paths of the failure mechanisms of interest
to this research and are based on the inclusion of the five failure factor groups described in
Section 4.3. They are utilised in the prototype system to help determine the most probable
(critical) failure path and the causes of failure. The predominant failure mechanisms are
included in the prototype system herein, namely rutting, fatigue cracking, and shear.
140
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Computational Model:
The support vector machines modelling technique was selected from those investigated in
Chapter Five as the most suitable for predicting the likelihood of road pavement failure from
Four processes are used to determine the likelihood of pavement failure from road datasets.
These are:
2. Model each combination of the failure factors, as per Table 5-4, to determine the
to identify from the available research dataset the combinations of factors which
could cause failure, based on an accuracy percentage of 100 % for rutting and
fatigue cracking failures and 90 % for shear failure. The probability (PFailure) for
rutting, fatigue cracking, and shear failure is calculated using Equation 6-1,
ி௨ = "#$
%,
&, … ,
'
Equation 6-1
ℎ =
ℎ , =
ℎ , =
ℎ
4. Assess the interactions between the failure mechanisms to calculate the overall
failure probability. For example, these calculations may take into account multiple
141
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Chapter Four identified the five failure factor groups contributing to failure as traffic,
composition, strength, environment, and subgrade sensitivity, which were considered in the
Chapter Five identified the successful factor combinations using the support vector machines
technique, as shown in Table 6-1. Each combination was superimposed on the developed
failure charts, as shown in Figures 6-2, 6-3, and 6-4, to represent each possible failure path as
identified by Table 5-3. From herein, the term ‘failure path’ refers to these combinations. The
combinations marked with an asterisk in Table 6-2 do not feature on the charts presented here
as they may be either a combination of other failure paths or are not significant causal factors.
Table 6-1: Successful Factor Combinations for the Support Vector Machines Technique
142
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Table 6-2: Factor Combinations of the State Highway LTPP Dataset per Failure Mechanism
Fatigue
Trial Combinations of Factors Rutting Shear
Cracking
3 Strength *Yes *Yes -
7 Traffic + Composition Yes Yes Yes
8 Traffic + Strength Yes Yes -
12 Composition + Strength Yes Yes Yes
16 Strength + Environment *Yes - -
18 Strength + Subgrade Sensitivity Yes Yes Yes
22 Traffic + Composition + Strength Yes Yes Yes
23 Traffic + Composition + Environment - Yes -
24 Traffic + Composition + Surface Condition - Yes -
25 Traffic + Composition + Subgrade Sensitivity Yes Yes Yes
26 Traffic + Strength + Environment Yes Yes -
28 Traffic + Strength + Subgrade Sensitivity Yes Yes Yes
32 Composition + Strength + Environment Yes Yes Yes
33 Composition + Strength + Surface Condition - Yes -
34 Composition + Strength + Subgrade Sensitivity Yes Yes Yes
39 Strength + Environment + Subgrade Sensitivity Yes - -
42 Traffic + Composition + Strength + Environment Yes Yes Yes
43 Traffic + Composition + Strength + Surface Condition - Yes -
44 Traffic + Composition + Strength + Subgrade Sensitivity Yes Yes *Yes
46 Traffic + Composition + Environment + Subgrade Sensitivity Yes Yes -
Traffic + Composition + Surface Condition + Subgrade
47 - Yes -
Sensitivity
49 Traffic + Strength + Environment + Subgrade Sensitivity Yes Yes Yes
50 Traffic + Strength + Surface Condition + Subgrade Sensitivity - Yes -
52 Composition + Strength + Environment + Surface Condition - Yes -
53 Composition + Strength + Environment + Subgrade Sensitivity Yes Yes Yes
Composition + Strength + Surface Condition + Subgrade
54 - Yes -
Sensitivity
Traffic + Composition + Strength + Environment + Surface
57 - Yes -
Condition
Traffic + Composition + Strength + Environment + Subgrade
58 *Yes *Yes *Yes
Sensitivity
Traffic + Composition + Strength + Surface Condition +
59 - Yes -
Subgrade Sensitivity
Composition + Strength + Environment + Surface Condition +
62 - Yes -
Subgrade Sensitivity
Traffic + Composition + Strength + Environment + Surface
63 - *Yes -
Condition + Subgrade Sensitivity
TOTAL 19 38 13
- Not applicable
* Not shown on the failure charts
143
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
144
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
145
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Data from the State Highway LTPP road network (refer Section 3.3.1) was used to further
develop and demonstrate the performance of the prototype system. The support vector
machines computational model was trained using 90 % of the data, via a 10-fold cross-
validation approach (refer Section 5.3.3) (Rogers and Girolami, 2012). Accordingly, the
calculations of the performance measures of the prototype system (see Section 6.4) used the
predictions (new observations) from the remaining 10 % of the dataset. However, the analysis
146
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
of the overall failure probability of the State Highway LTPP road network was based on the
The prototype system was evaluated using the performance measures discussed in Chapter
Five (refer Section 5.3.2). To this end, the trained models were evaluated by comparing the
failed sites predicted by the support vector machines technique against the number recorded
to have failed in the dataset. This was carried out on a failure path basis and the results are
shown in Tables 6-3, 6-4, and 6-5. For example, it may be seen from Table 6-3 below that the
trained model reported an accuracy of 98 % in predicting rutting failure based on failure path
7. As the system modelled each failure path separately, the performance of the system is
Tables 6-3, 6-4, and 6-5 show the results the performance analysis of the system in terms of
the four measures used. From the results, it may be seen that the system is able to predict the
occurrence of rutting and fatigue cracking to a high degree of accuracy (with an average of
approximately 98 % in both cases), but less so shear failure (94 %), similar to the poor results
presented in Chapter Five for predicting shear failure. This suggests the independent
variables do not correlate well with the predictions, such that the dataset does not contain
appropriate information for predicting shear failures. The F-Score (99 % for both rutting and
fatigue cracking, and 97 % for shear) indicated the superior accuracy and effectiveness of the
system to predicting failure. Overall, the results show consistency for each of the factor
combinations in terms of the accuracy, misclassification error, and F-score measures, which
147
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
demonstrates the success of informing the model with the engineering knowledge captured in
the failure charts. However, the phi coefficient (average of 0.17, 0.28, and 0.13 for rutting,
fatigue cracking, and shear respectively) indicated a weak correlation between the inputs of
the system and the predicted failure probabilities, despite the success of the support vector
machines technique reported in Figure 5-6. However, in each table, a number of failure paths
improve on this average, suggesting that the fatigue cracking and rutting systems in particular
sufficiently model the behaviour of pavement failure. The poor result of the shear failure
system, similar to that reported earlier in Figure 5-6, further substantiates the reported
influence of the aggregate behaviour on shear failures in the literature, as material types were
excluded in the research dataset. A phi coefficient equal to zero (0) indicates no correlation
between the inputs and predicted output. Although this result can also occur when one of the
binary classes is not predicted in any of the 10 cross-validation tests, resulting in the sum of a
predicted class is equal to zero (0), the results reported in the tables below list the phi
coefficients which were unable to be calculated with a ‘-’ to avoid any confusion.
The coefficient of variation was calculated using Equation 6-2 (Madsen, 2011) allowing a
direct comparison between populations with different means; therefore, the four performance
measures across the three failure mechanisms can be compared. In the tables below, the
accuracy and F-score measures are less dispersed than the misclassification error measure.
)// , I
Equation 6-2
! =
5
× 100 = × 100
148
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Failure Misclassification
Accuracy (%) F-Score Phi Coefficient
Paths Error (%)
3 97.651 2.349 0.988 -
7 97.865 2.135 0.989 0.341
8 97.651 2.349 0.988 -
12 97.794 2.206 0.989 0.301
16 97.669 2.331 0.988 0.114
18 97.651 2.349 0.988 -
22 97.794 2.206 0.989 0.301
25 97.794 2.206 0.989 0.301
26 97.669 2.331 0.988 0.128
28 97.651 2.349 0.988 -
32 97.633 2.367 0.988 0.163
34 97.794 2.206 0.989 0.301
39 97.705 2.295 0.988 0.153
42 97.651 2.349 0.988 0.168
44 97.794 2.206 0.989 0.301
46 97.633 2.367 0.988 0.210
49 97.722 2.278 0.988 0.181
53 97.616 2.384 0.988 0.158
58 97.633 2.367 0.988 0.163
Minimum 97.616 2.135 0.988 0.114
Average 97.704 2.296 0.988 0.219
Maximum 97.865 2.384 0.989 0.341
Coefficient
0.5 20.1 0.3 15.71
of Variation
1
– Calculated with the ‘-ʼ omitted
149
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Failure Misclassification
Accuracy (%) F-Score Phi Coefficient
Paths Error (%)
7 94.520 5.480 0.972 0.186
12 94.431 5.569 0.971 0.149
18 94.448 5.552 0.971 -
22 94.555 5.445 0.972 0.196
25 94.626 5.374 0.972 0.202
28 94.448 5.552 0.971 -
32 94.520 5.480 0.972 0.138
34 94.591 5.409 0.972 0.188
42 94.502 5.498 0.972 0.133
44 94.662 5.338 0.972 0.214
49 94.413 5.587 0.971 0.053
53 94.520 5.480 0.972 0.133
58 94.520 5.480 0.972 0.138
Minimum 94.413 5.338 0.971 0.053
Average 94.520 5.480 0.972 0.157
Maximum 94.662 5.587 0.972 0.214
Coefficient
0.5 7.9 0.2 -1
of Variation
1
– Cannot be calculated due to a negative standard deviation (see Appendix F)
150
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Table 6-5: Assessment of the Performance Measures for Fatigue Cracking Failure
(Average of 10-Fold Cross-Validation)
Failure Paths Accuracy (%) Misclassification Error (%) F-Score Phi Coefficient
3 98.078 1.922 0.990 -
7 98.274 1.726 0.991 0.373
8 98.078 1.922 0.990 -
12 98.274 1.726 0.991 0.373
18 98.078 1.922 0.990 -
22 98.220 1.779 0.991 0.358
23 98.292 1.708 0.991 0.363
24b 98.149 1.851 0.991 0.261
25 98.220 1.779 0.991 0.358
26 98.007 1.993 0.990 0.063
28 98.078 1.922 0.990 -
32 98.381 1.619 0.992 0.406
33a 98.185 1.815 0.991 0.275
33b 98.149 1.851 0.991 0.261
34 98.274 1.726 0.991 0.373
42 98.381 1.619 0.992 0.417
43a 98.221 1.779 0.991 0.296
43b 98.185 1.815 0.991 0.282
44 98.221 1.779 0.991 0.383
46 98.292 1.708 0.991 0.374
47b 98.167 1.833 0.991 0.275
49 98.007 1.993 0.990 0.063
50a 98.185 1.815 0.991 0.267
50b 98.149 1.851 0.991 0.253
52a 98.221 1.779 0.991 0.303
52b 98.256 1.744 0.992 0.334
53 98.452 1.548 0.992 0.447
54a 98.149 1.851 0.991 0.253
54b 98.149 1.851 0.991 0.261
57a 98.221 1.779 0.991 0.303
57b 98.274 1.726 0.991 0.346
58 98.381 1.619 0.992 0.417
59a 98.221 1.779 0.991 0.296
59b 98.167 1.833 0.991 0.275
62a 98.203 1.797 0.991 0.296
62b 98.238 1.762 0.991 0.328
63a 98.221 1.779 0.991 0.303
63b 98.256 1.744 0.991 0.334
Minimum 98.007 1.548 0.990 0.063
Average 98.209 1.791 0.991 0.311
Maximum 98.452 1.993 0.992 0.447
Coefficient of
0.6 35.5 0.3 1.01
Variation
1
– Calculated with the ‘-ʼ omitted
151
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
The development of the computational model described above and in Chapter Five has
assumed that each of the three failure mechanisms act independently. However, in practice,
another occurrence of failure. For example, rutting can act alone on the pavement or it can
permit water into the lower layers of the pavement resulting in rutting. To address this
phenomenon in the computational model, this research explored probability theory to account
6.5.1 Terminology
A number of terms are used to describe inter and independence of events in probability
theory (Mendenhall and Beaver, 1991). Definitions of interest to the problem described above
are:
Mutually Exclusive: A mutually exclusive event is one in which two events cannot
happen at the same time (Ayyub and McCuen, 2003; Mendenhall and Beaver, 1991),
as shown in Figure 6-5. The fact that any of the three defects considered herein can be
present on a road section with any of the other defects means that their occurrence is
152
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
McCuen, 2003; Mendenhall and Beaver, 1991). For example, cracking occurring as a
secondary effect of rutting would indicate these events are dependent (not
Probability Unions: The ‘Additive Law of Probability’ (Ayyub and McCuen, 2003;
Mendenhall and Beaver, 1991) defines the probability of the occurrence of two events
as:
(% ∪ &) =
% +
& −
% ∩ &
Equation 6-3
153
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
(% ∪ &) =
% +
&
Equation 6-4
If events are defined as independent, Equation 6-3 becomes (Ayyub and McCuen,
% ∪ & =
% +
& −
% ∩ &
Equation 6-5
=
% +
& − $
% × (&)'
Given the complexity of the problem of calculating the overall failure probability, three
the maximum probability of each of the three failure mechanisms (refer Equation
3-1);
Equation 6-3, assuming the failure mechanisms are independent of each other
154
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
With this conservative approach, the maximum of the probabilities from three individual
failure mechanisms is taken as the overall failure probability (PFailure) of the road section
Equation 6-6
ி௨ = "#Jோ௨௧௧ , , ௌ K
Advantages: This approach follows that of conventional pavement design in which the
design of the pavement structure focuses on the critical failure mechanism(s), whether that is
for example to prevent rutting or cracking (Austroads, 2012). This approach treats each
Limitations: This approach assumes that each failure mechanism is independent of the others
and, therefore, evidence of the likelihood of two or more failure mechanisms on a road
The laws of addition of probabilities can be used to calculate the probability of failure by any
one, two or three of the failure mechanisms acting concurrently. (refer Section 3.2.4). The
law can be written using ‘set’ notation as follows (Ayyub and McCuen, 2003; Mendenhall
ி௨ =
% +
& +
−
% ∩ & −
% ∩ −
& ∩ +
% ∩ & ∩
Equation 6-7
−
6 ∩ )ℎ −
!6 ∩ )ℎ
+
6 ∩ !6 ∩ )ℎ
155
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Note if one failure mode is considered not to influence the occurrence of the other (e.g. the
be written as:
Equation 6-8
ி௨ = ோ௨௧௧ + + ௌ − Jோ௨௧௧ × K − Jோ௨௧௧ × ௌ K
Advantages: This approach takes into account the possibility of multiple failures occurring.
Limitations: Although this approach accounts for multiple failures in the calculations, there
is not enough information in the dataset to determine the timings of failure. More specifically,
it is not possible to determine from the dataset whether the failure modes that occurred
simultaneously were in fact secondary effects of each other (e.g. one mode influenced the
The overall failure probability could be determined using a modelling approach to predict the
occurrence of combined failures. To achieve this, individual models would be developed for
each combined failure occurrence, much like the individual failure paths in the failure charts
(refer the methodology presented in Section 6.4). The model outputs would then predict the
individual probabilities for each failure type and combined failures, as shown in Equation 6-
9.
ி௨ =
ାௌ , ோ௨௧௧ାାௌ
Advantages: This approach presents a superior method of calculating the overall failure
probability.
156
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
train the computational model. As the State Highway LTPP dataset contains only a small
number of multiple failure occurrences (see Table 6-6) and limited information on the
occurrences of multiple failure modes, it was not possible to develop such a process in this
research.
The above section explored three approaches to determine an overall failure probability for a
road section. The third approach was found to be difficult to achieve using the available
dataset and therefore will not be considered further. For the purposes of the prototype system,
it was assumed the failure mechanisms were independent, such that one mode of failure has
no influence on the occurrence of another. For a more accurate assessment, the dataset would
need to include more information regarding the timing of failure types and the independence
between failure mechanisms. Consequently, approaches one and two described above were
further explored using the State Highway LTPP dataset, as described below.
157
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
In order to compare the results obtained from the two approaches described above,
distribution plots of the overall failure probabilities across the State Highway LTPP road
network were generated and compared. The performance measures, defined in Section 5.3.2,
Figure 6-7 presents the failure probability distribution of the network using the conservative
(maximum probability) approach to calculate the overall failure. By inspection, the actual
failure data, where the failure variable was converted to an estimate of the probability density
function, of this network indicated a similar failure distribution, as shown by the black
density curve in Figure 6-7. A large number of sites in the State Highway LTPP dataset have
not failed, hence the majority of the distribution has a failure probability of less than 20 %.
Figure 6-7: Distribution of the overall failure probabilities, calculated using the maximum probabilities,
of the State Highway LTPP road network
158
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Figure 6-8 presents the distribution of failure probabilities determined using the second
(probability theory) approach. In comparison to Figure 6-7, although the peak of the
distribution is smaller than obtained by the first approach, the shape of the distribution
remains the same with a slight shift to the right suggesting that more sites have a higher
likelihood of failure. This result could be expected given the addition of the individual failure
Figure 6-8: Distribution of the overall failure probabilities, calculated using the probability equation, of
the State Highway LTPP road network
Figure 6-9 directly compares the distributions of the two approaches, where the differences in
the distributions, such that the overall failure probability calculated using the probability
theory (probability equation) approach would always exceed that of the conservative
159
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
To further compare the results from the two approaches, confusion matrices (refer Chapter
Five) were generated, as shown in Figure 6-10, based on the outputs using a failure threshold
of
; ≥ 0.5 = 1. The results for the two approaches are very similar, with the only
difference being the very small increase (0.107 %) in the number of misclassified non-
PREDICTIONS
0 1 0 1
1 54 77 131 1 60 77 137
5280 348 5628 5280 348 5628
(a) from the Conservative Approach (b) from the Probability Theory Approach
(Maximum Probability) (Probability Equation)
Figure 6-10: Confusion matrix of the results
160
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
The data in Figure 6-10 was further analysed; the results from the performance measures for
each approach is summarised in Table 6-7. The results of these analyses show that in terms of
the performance measures considered there is an insignificant difference between the two
approaches and that either approach may be suitable for the dataset considered.
Table 6-7: Performance Measures for the Conservative and Probability Theory Approaches
Misclassification F- Phi
Accuracy
Error Score Coefficient
Conservative Approach
94.23 % 5.77 % 0.97 0.34
(Maximum Probability)
Probability Theory Approach
94.12 % 5.88 % 0.97 0.33
(Probability Equation)
This chapter has presented a prototype system which includes failure charts (developed in
Chapter Four) and a computational model (selected in Chapter Five). The system comprises
• Modelling each possible combination of the factor groups to select the successful
combinations;
and
• Assessing the approaches to determine the overall failure probability using the
predicted probability outputs from the prototype system for each failure
mechanism.
161
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
Chapter Five identified the successful factor combinations and this chapter presented them on
the developed failure charts (from Chapter Four). Each combination represents a failure path
for each failure mechanism. Not all the same combinations are present on each failure chart,
as the mechanisms associated with each failure are different. However, the successful factor
combinations demonstrate the key factor groups and failure paths appropriately for each
failure mechanism.
To calculate the failure probabilities of each failure mechanism, the probability associated
with the most probable failure path (e.g. the maximum probability for the failure mechanism)
was selected. With this approach, the factor combination associated with the greatest
probability becomes the critical failure path. Identifying this path and combination of factors
on the failure paths indicates the most likely causes of the failure.
This chapter assessed the performance of the prototype system based on the four performance
measures introduced in Chapter Five, from which it was found that the performance of the
To calculate the overall failure probability, three approaches were investigated and, of these,
two were selected for further analysis. The dataset did not contain information on the timings
approach.
The two approaches selected were used to produce the distributions of the overall failure
probabilities for the State Highway LTPP dataset. The first (conservative) approach selected
the maximum probability from the three failure mechanisms as the overall failure probability.
The second approach used the principles of probability unions to develop an equation to
162
Chapter Six: DEVELOPMENT OF THE PROTOTYPE SYSTEM
Predicting the Probability of Failure
The following chapter further assesses the performance of the developed system, given both
163
164
Chapter Seven
7.1 Introduction
To demonstrate the effectiveness of the prototype system and the practical applications of this
research outcome, as stated in Objective Four, this chapter presents the results of the analysis
of an independent dataset from the New Zealand LTPP programme. This case study aims to
• Calculate the probability of failure for road sections in the dataset, and determine
165
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
• Identify the symptomatic problems across the pavement sites included the dataset
• Critically analyse the shifts in the probability distributions to quantify the effect
the changes in the environment have on the road network, such as an increase in
The above includes both project and network level applications of the prototype system. With
such an analysis, asset managers can make informed decisions on the future maintenance
demands of their network, and identify potential failures. This chapter discusses the results
from the analysis illustrating the performance of the sites included in the Local Authority
To demonstrate the use of the system, the case study followed the methodology below.
2. Process the data using the support vector machines models developed in Chapter
Six (note: the models were not retrained using the testing dataset; instead the
4. Assess the accuracy of the predictions from the system using the predicted failure
5. Compare the overall failure probability using the two approaches discussed in
166
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
Statistical plots, such as histograms, barplots, and pie charts, were used to gain an
The Local Authority LTPP dataset (refer Section 3.3.2) was processed using the prototype
system developed in Chapter Six to further test the performance of the system on an
independent dataset, one which was not used to develop and refine the prototype system, and
evaluate the transferability of the system to other road networks. The similar data collection
methods employed under the LTPP programme (refer Section 3.3) consequently led to
similar variables included in the Local Authority LTPP dataset as those included in the State
Highway LTPP dataset, and subsequently the successful implementation of the prototype
system to another dataset. Alternative data collection methods are implemented by RCAs on
typical local authority road networks, such as the network analysed in Schlotjes and Henning
(2012), resulting in distinct differences in the variables included in the datasets and databases
in comparison to the research dataset. Thus, evaluating the performance of the prototype
system on such a network would vary the predictions for the pavements and ultimately
mislead and falsify the performance results, given the expected high percentage of missing
Testing of the system to real-life data, such as New Zealand RCA road networks, requires
manipulation of the network data to maximise the amount of data available for the prototype
system. Deng and Henning (2012) and Schlotjes and Henning (2012) tested the system with
Table 7-1 summarises the failure distribution of the Local Authority LTPP dataset by
geographical region. Similar percentages of failure were seen on the State Highway LTPP
167
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
road network (refer Table 6-6), although the discernible difference is a reduction of 1 % from
the State Highway LTPP dataset in rutting failures, and an increase of 1.1 % for the Local
Authority LTPP dataset in fatigue cracking failures. The majority of the road sites in the State
Highway LTPP dataset, which was employed in the development of the prototype system,
were located in rural environments. Therefore, to ensure consistent and indicative results
from the system testing, this case study disregarded any urban sites included in the Local
Authority LTPP dataset as the probabilities of these sites were not able to be appropriately
predicted by the prototype system. Although the case study only concerned the rural LTPP
sites and, as a result, reduced the number of datapoints included in the independent testing
dataset, the rural sites provided 4136 datapoints for the testing dataset. The size of this testing
dataset was similar to the size of the development dataset (State Highway LTPP dataset),
which contained 4512 datapoints, and therefore can be considered sufficient for the purpose
of this research.
The rural sites included in the LTPP programme are situated in several geographical regions
of New Zealand. Generally, the traffic loadings on these sites is much less than the demand
on the State Highway road network, resulting in a thinner pavement design and often lacking
a sub-base layer. Despite this, a number of less popular State Highway routes are similar in
design, construction, traffic and environmental conditions, and regional locations as the Local
Authority road pavements, resulting in the verification of the prototype system using this
dataset feasible.
168
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
Table 7-2 presents a comparison between the recorded failures given in the rural Local
Authority LTPP dataset and those predicted by the prototype system. Overall, the system was
less accurate at predicting failure for the rural Local Authority LTPP dataset than that used
for the development of the prototype system in Chapter Six. This was expected as using the
system to analyse another road dataset, although of the same detail and including similar
pavement types, was expected to bring about a reduction in performance as it is common for
computational models to lose their predictive power and robustness through validation of the
suggested further refinement of the system is required, which may include adequately
redefining the developed failure charts when the system is used on other datasets and
pavement types.
Unlike the results reported in Chapter Six, Table 7-2 shows a similarity between the accuracy
of the rutting and shear failure systems. This result was unexpected and suggests that the
failure mechanisms described in Figure 4-7, and subsequently the prototype system, may not
include all methods of rutting failure included in this particular testing dataset. This also
demonstrates the need to adequately redefine the developed failure charts when the system is
169
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
used on other datasets. Furthermore, it suggests the performance of the rutting system is
compromised, more so than the other two, when transferring the developed support vector
machines models to other road datasets. Despite this, the performance of the prototype
system, and subsequently the failure probability predictions, were considered adequate for the
testing dataset, such that the transferability of the prototype system to another road dataset
was successful.
The project level applications of the proposed system include the ability to:
Using these outputs in conjunction with the failure charts developed in Chapter Four
facilitates an appropriate diagnosis of the cause(s) of failure and the selection of suitable
maintenance treatment(s) for a road section. These aspects are discussed further in
170
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
In order to demonstrate the project level applications of the prototype system, 12 sites were
selected at random15 from the entire rural Local Authority LTPP dataset, as presented in
Table 7-3. The predicted failure state of each road section was predicted using the developed
prototype system and Equation 6-6. Sections which have been predicted to have failed are
indicated with a 1 in the ‘Failure’ column. Then, for the purpose of this case study, the
overall probability of failure was calculated using the conservative (maximum probability)
approach (refer Section 6.5.2.1). The most probable failure mode and, subsequently, the
associated most probable failure path are also presented in Table 7-3. In the instances where
failure was not predicted to occur, the identified failure mode indicates the mechanism by
15
Site numbers were randomly generated by the random number generator sample function in R.
random_sites <- sample ( 1 : 4136 , 12 , replace = F )
171
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
Table 7-3: Inferred Project Level Statistics from the Rural Local Authority LTPP Dataset
To further demonstrate the use of the system, Table 7-4 presents some additional information
on secondary failure predictions for the 12 sites, which has been determined using the second
16
Despite the site numbers exceeding 4136, the sites reported in Table 7-3 are rural sites only. The urban sites
are removed from the dataset yet the original site numbers remain from the entire Local Authority LTPP dataset.
172
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
Table 7-4: Secondary Failure Predictions of the Rural Local Authority LTPP Dataset
Two sites are further explained to demonstrate the project level applications. Site # 512, as an
example, is likely to fail in rutting with a probability of 0.875, which correlates well with the
raw data. The predicted critical failure path for this site is # 22 (refer Table 7-3), indicating
that a combination of the traffic, poor composition, and weak pavement strength are the
contributing factors to the rutting failure. Figure 7-1 shows the use of the failure charts for
this task, and with reference to the developed rutting failure chart presented in Chapter Four
(Figure 4-7) and again in Chapter Six (Figure 6-2), failure path # 22 is defined as:
173
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
resulting in pavement layers that are too thin, or the use of poor aggregates, which are unable
to spread the traffic load sufficiently to the underlying layers. A suitable maintenance
reduce the strains adequately in the underlying layers for the traffic loadings, thus preventing
rutting from occurring in the future. The required thickness could be calculated using an
Site # 4962 failed in shear, and the prototype system predicted shear failure as the most
probable failure mode with an associated probability of 0.506 and a predicted secondary
failure of rutting with a probability of 0.347 (refer Table 7-4). The system identified the
critical (shear) failure path as # 42 (refer Figures 4-11 and 6-4) and Figure 7-2 identifies the
174
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
Materials → Environment”
Failure path # 42 includes the environmental factor group indicating this shear failure can be
attributed to water ingress in the structural layers of the pavement, which in turn weakens the
materials. Under repetitive traffic loads, the pavement results in material shear of the
basecourse layer. Cracks in the surface layer or a sub-terrain drainage problem are the likely
causes of the water ingress in this pavement. Therefore, the appropriate maintenance
treatment would be to seal the cracks with a resealing exercise and / or remedy the drainage
failure path # 49) to prevent the occurrence of the predicted secondary failure. Based on
Figure 6-2, secondary rutting is likely to occur due to a weakened subgrade as a result of
water ingress, which coincides with the above problem resulting in shear failure. Therefore,
175
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
addressing the ingress of water into this pavement structure will likely prevent both primary
At a network level, the developed prototype system can assess the performance of the road
• The proportion of the network susceptible to failure and the distribution of the
• The distribution of the most probable causes of failure (critical failure paths) of
problems.
Figure 7-3 presents the distribution of the overall failure probabilities, based on the
distribution of the predicted probabilities is similar to that of the actual failure distribution
(see Figure 7-3) where the peaks of the distribution curve occur in similar probability ranges
of the failure spectrum, despite the peak at the lower end of the probability scale having
shifted slightly to the right. Approximately one-third of the rural sites included in the rural
Local Authority LTPP dataset are predicted to fail, given sites with a predicted probability of
overall failure greater than 0.5. However, shifting the failure threshold to a higher value (0.8)
176
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
resulted in approximately 20 % of the sites being predicted to fail. Although the predicted
failure curve has shifted to the right of the actual failure density curve, the predicted failure
distribution presents a large proportion of road sections with low probability of failure
(probability less than 0.2) and a secondary peak with a higher failure probability (greater than
0.95).
3.0
Non-Failure Failure
Number of Predicted Non-Failures = 2961 Number of Predicted Failures = 1175
Non-Failure Failure
Number of Predicted Non-Failures = 3291 Number of Predicted Failures = 845
Density Curves
2.5
Predictions
Actual
1000
2.0
Frequency
Density
1.5
500
1.0
0.5
0.0
0
Failure Probability
Figure 7-3: Failure profile of the rural sites of the Local Authority LTPP dataset, using the conservative
(maximum probability) approach
The second approach to calculating the overall failure probability, using additive probability
theory (refer Section 6.5.2.2), also assessed the performance of the rural Local Authority
LTPP dataset. Again, Figure 7-4 reported similarities between the predicted and actual failure
distributions; however, the peak centred around 0.15 is noticeably greater than that previously
reported in Figure 7-3. With this approach, a higher percentage (36 %) of the road sections
177
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
are predicted to fail (see Figure 7-4), reducing to 22 % when the failure threshold is increased
to 0.8.
3.0
Non-Failure Failure
Number of Predicted Non-Failures = 2634 Number of Predicted Failures = 1502
Non-Failure Failure
Number of Predicted Non-Failures = 3233 Number of Predicted Failures = 903
Density Curves
2.5
Predictions
Actual
1000
2.0
Frequency
Density
1.5
500
1.0
0.5
0.0
0
Failure Probability
Figure 7-4: Failure profile of the rural sites of the Local Authority LTPP dataset, using the probability
theory (probability equation) approach
Although the differences and shifts between Figures 7-3 and 7-4 are a result of two methods
of calculating the overall failure probability, the figures provide an example of identifying the
changes in the failure profiles. For example, given a change in the road network’s
environment, such as an increase in traffic loadings coupled with environmental changes and
/ or increased rainfall, the data for the modified (increased traffic) scenario can be used in the
prototype system. The predicted probabilities for both the original and modified scenario can
then be analysed visually to attain the shifts in the distribution profiles, similar to the above
Figures 7-3 and 7-4, providing valuable information to the asset manager. An application of
this approach would be useful in cases for example where a re-routing exercise increases the
178
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
traffic loadings on a less strategic road, which typically had not been designed to carry the
increased traffic loading levels. Schlotjes and Henning (2012) reported on this application of
the developed prototype system using a typical road network in New Zealand, which recently
The most probable failure modes predicted by the prototype system were used to determine
the susceptibility of the rural Local Authority LTPP road pavements to specific failure types.
The two approaches, to calculate the overall failure probability described in Chapter Six,
were used to determine the susceptibility of the network to failure types comparatively, as
shown in Figure 7-5. From the distribution of failure modes associated with the predicted
failed road sections, the majority of the network is predicted to fail in rutting, as expected.
However, fewer sites are predicted to fail in rutting using the probability theory (probability
equation) approach (refer Section 6.5.2.2) and a different failure mode distribution is
presented.
When the failure threshold is increased from 0.5 to 0.8 for both approaches, the proportion of
rutting failures also increased, suggesting a susceptibility of this network to rutting. The
fatigue cracking proportions remain unchanged, yet the shear failures at the higher end of
failure probability distribution decreased, in some cases by half. The apparent difference
179
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
Figure 7-5: Predicted failure modes of the rural sites of the Local Authority LTPP dataset
Analysing the predicted critical failure paths identifies the symptomatic problems across the
application, Figures 7-6, 7-7, and 7-8 present the frequencies of occurrence of the critical
Figure 7-6 shows little difference between the two approaches used to determine the
probability of failure, for the predicted failure paths of rutting failures. The predominant
causes of network-wide rutting were traffic loadings, pavement composition and pavement
strength (failure path # 22). From Figure 7-8, weaknesses in the pavement composition
180
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
coupled with traffic loadings (failure path # 7) caused the majority (44 % and 34 %) of
network-wide fatigue cracking, given the respective approaches of calculating the overall
failure. However, two distinct failure paths were identified in Figure 7-7; therefore, the
predominant causes of shear include traffic loadings, pavement composition, and pavement
strength (failure path # 22), and the above (failure path # 22) in conjunction with subgrade
181
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
182
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
Figure 7-8: Critical failure paths for predicted fatigue cracking failure
183
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
This chapter analysed the rural Local Authority LTPP dataset to demonstrate the practical
applications, both at project level and network level, of the developed prototype system and
evaluate the performance of the system on an independent dataset. The system performed
adequately on the independent testing dataset; however, the performance of the rutting
system in comparison to Chapter Six was inferior. This may be attributable to the system not
recognising or including all methods of rutting failure, which were present in the Local
Authority LTPP dataset (testing dataset), and suggests the need to refine the rutting system
further.
From the network dataset, 12 sites were selected at random to demonstrate the use of the
system. For each, the probability of failure was reported and the most probable causes of
failure were identified. This information can assist the asset manager in maintenance
decisions.
At a network level, Figures 7-3 and 7-4 showed the distributions of the predicted failure
probabilities followed that of the actual failures presented in the rural Local Authority LTPP
dataset and a large portion of the network were not predicted to fail ((; < 0.5)). The
primary failure mode of the rural Local Authority LTPP road pavements was rutting. Such
Overall, the developed prototype system can be employed in the management of road
networks to:
184
Chapter Seven: NEW ZEALAND CASE STUDY
Practical Applications of the System
The results of the prototype system may be regarded as promising for the potential of an
effective non-destructive diagnostic tool. However, further testing of the developed failure
system is both desired and necessary before it can be considered to be fully developed. This
should include further testing of the performance of the system on network data. The
following chapter summarises the reported research and reviews the methodology and
185
186
Chapter Eight
8.1 Introduction
The development of a prototype system that predicts the probability and diagnoses road
pavement failure is a novel approach to supplement the current PMS and road reporting
processes. Current practices fail to present a robust methodology for predicting the failure
probability of road pavements for further use in network assessments of failure risk. To
address this, the framework developed in this research provides a means to predict the
probability of failure of road sections and to identify the most probable causes of failure.
The purpose of this chapter is to provide a critical review of the methodology followed in this
research. The research objectives outlined in Chapter One were accomplished by:
187
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
The research aim, defined in Chapter One, was to predict the probability of road pavement
failure. In the LTPP datasets, failure was represented with a binary class label, such that a
road section was defined either as a failed site or a sound site; thus, mathematically defining
the research task as a binary classification problem, focussing on the end of life state of the
pavement as opposed to the gradual deterioration of pavements over time. Although other
deterioration models, the definition of the research task limited the modelling approaches
188
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
Based on the findings in the literature review (Chapter Two), a methodology was proposed to
develop a classification model, selected through the comparative study, which quantified the
probability of road pavement failure. Through the analysis of the system, it was concluded
that the methodology followed in this research successfully inferred engineering knowledge
1. Develop charts for rutting, cracking, and shear failures that supplements the
current knowledge base and identifies the factors contributing to failure and
the State Highway LTPP dataset, based on a developed selection criteria process
and a review of suitable classification techniques. This study aimed to evaluate the
predicted outputs from the system included a probability of road pavement failure
189
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
Failure charts illustrating the causes of rutting, cracking, and shear failures were
developed.
Chapter Four reviewed the fundamental knowledge of each failure mechanism, canvassed
expert opinion on road pavement failure, and explored potential relationships in two New
Zealand road datasets. The developed failure charts captured this engineering knowledge of
pavement failure by presenting the possible causes and illustrated the failure paths associated
with each failure mechanism. Five failure factor groups were inferred from the developed
failure charts for further use in modelling processes, as described in Chapter Five. The charts
were revisited in Chapter Seven and used to identify the most probable cause(s) of failure.
In addition to the application demonstrated in this thesis, the failure charts can provide asset
managers, civil engineers, and RCAs a comprehensive background to the causes of rutting,
techniques.
The reviewed literature revealed several classification techniques that performed well in the
respective study domains. Five of these were selected to assess their suitability for the use
with road pavement data using a number of criteria, including the advantages and limitations
of each approach. Four point performance measures evaluated the performance of each
modelling technique. Hypothesis testing was employed to determine if the difference between
the results of the performance measures, for each modelling technique, was statistically
190
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
significant or not, and to conclude whether one modelling technique outperformed the others.
Based on a further set of criteria including model performance, user interpretability, and
computer efficiency, the performance and effectiveness of the five classification techniques
were evaluated. This demonstrated that the support vector machines technique was most
suitable for the task at hand and accordingly it was selected to develop further into a
prototype system.
failure charts into computational models. Each failure mechanism was modelled individually
and the resulting failure probability for each mode of failure was assumed to be the maximum
probability across all failure paths modelled. Three approaches were considered to calculate
the overall failure probability; however, the analysis failed to conclude on one superior
approach. Only two approaches were pertinent to this research, namely the conservative
(maximum probability) approach (refer Section 6.5.2.1) and an approach based on probability
theory (refer Section 6.5.2.2). The proposed computational modelling approach (refer Section
6.5.2.3) was thought to be the best calculation of the overall failure probability; however,
given the few occurrences of multiple failures in the research dataset, the implementation of
The effectiveness of the prototype system was demonstrated using the Local Authority
LTPP dataset.
Chapter Seven presented the results from the system testing on an independent New Zealand
road network. The accuracy of the system was found to be sufficient and demonstrated the
191
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
transferability of the system to a new dataset. The chapter further discussed the practical
• Predicting the probability of road pavement failure for both overall and the
• Assessing the impact of changes in the road environment and evaluating the
The methodology adopted for the development of a system, which quantifies the probability
of pavement failure, focussed on the end of life probability state of the pavement as opposed
methodologies adopted in this research could have been followed, which are explored and
1. Failure knowledge;
5. Data issues.
The developed failure charts made use of expert opinion, the literature, and knowledge
obtained from scrutinising road datasets in presenting the causes of rutting, cracking, and
192
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
shear failures of flexible road pavements. Adopting the robust methodology resulted in a
comprehensive understanding of road pavement failure obtained from the three sources of
knowledge. However, the causes included in these charts were site-specific to New Zealand
road environments; thus, for other road environments, amendments to the failure charts may
Inferring such engineering knowledge into the computational model identified the limitation
of assuming independence between the failure mechanisms in the development of the failure
charts, such that the support vector machines computational models developed for each
failure mechanism did not recognise the interactions between rutting, fatigue cracking, and
shear failures in the probability calculations. Rather, this aspect was addressed by using
probability theory (refer Chapter Six). To address this aspect, it is recommended that any
future work, which further refines the system developed herein, should focus on further
developing the failure charts so that they are able to depict possible interactions between
failure mechanisms.
Chapter Five presented a comparative study of five classification techniques using road
pavement data, as the performance of a classifier is highly dependent on the particulars of the
research dataset, thus the direct transfer of a single algorithm to another research domain is
uncertain (Dreiseitl and Ohno-Machado, 2002; Perlich et al., 2003). Other comparative
studies adopted techniques based on the popularity of the technique in the respective study
fields. This research developed objective-based criteria to select the techniques most suitable
for the research dataset, based on the performance of the techniques, user interpretability, and
193
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
However, the exclusion of other techniques at this stage of the research was reliant on the
A number of assumptions were adopted for the comparative study (Chapter Five) as
described below:
non-failures;
this research assumed it was equally undesirable for both incorrect predicted
It is inevitable for road network data to include missing values, as a result of financial,
equipment, and time restrictions imposed on the RCAs, such that inventory data in the
network databases is limited. For the State Highway LTPP dataset, missing values were
assigned a nominal value of zero (0) (refer Section 5.1.2.1), opposed to omitting such road
sections which would ultimately reduce the available data significantly. In many of the input
fields, this assumption was considered acceptable where it resulted in the worst possible
scenario of pavement composition and strength. For example, much of the State Highway
LTPP dataset failed to provide evidence of a sub-base layer and, in these cases, the thickness
and age of this layer was assigned a zero (0), indicating a thinner pavement. With this
approach, the predictions will be based off a pavement more susceptible to failure; this could
194
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
in turn distort the predictions of the new observations. For example, the model may assume a
thin pavement is sound, based on the data, but in fact it is a thicker pavement with an
unknown thickness of one of the layers that is sound. To address this, the data needs to be
assuming a zero (0) traffic loading represents the best possible scenario suggesting the
suggested adjusting the missing values using the mean of the input variables. For the State
Highway LTPP dataset, this approach would distort the representative sample of the network,
indicating the pavements were thicker and stronger than in reality, although for the
environment, traffic loadings, and condition data, the averages would represent the pavement
conditions realistically.
By adopting this assumption, the majority of the network assumed the worst possible
scenario, as the State Highway LTPP datasets included complete records of traffic loadings
and condition data. In the case of rainfall, incomplete data was cross-referenced to external
sources for completion. However, for road pavements, adopting this approach is recognised
to be misrepresentative of the data, and it is suggested to either adopt a mean nominal value
per input variable or a known default value appropriate for the respective network. Further
testing of the prototype system employed the latter to network datasets (Deng and Henning,
2012), who reported sufficient results from the road networks applicable to the scope of this
research.
Normalisation of the research dataset eliminates any preference from the model towards one
input variable over another, such that the amount of data considered by the model is
maximised. The performance of classifiers is improved when the input values range between
195
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
zero (0) and one (1) (Roiger and Geatz, 2003), typically achieved by the following
normalisation processes:
( =
"#" " − "" "
of the variable, such as the mean and standard deviation, are used to compute the
( =
// /
Although road pavement data seldom follows a normal distribution, the straight-line
assumed sufficient for this study, given the properties of the research dataset. As discussed in
Section 5.1.2.1, the uncertainty of the pavement behaviour and variability in road network
data distorts the distributions of the variables (Reigle, 2000), such that the distributions of the
variables are heavily dependent on the contents of the research dataset(s). For future adoption
of the framework, this research suggests investigating the distributions of the possible model
inputs to understand the nature of the variables and, subsequently, establish the appropriate
196
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
Classifiers perform best with a balanced dataset that has equal representation of classes, in
order for the model to assume the equal possibility of predicting a class for the new
observations. With unbalanced data classes, the trained model will always attempt to
minimise the error, thus the majority of the new predictions will belong to the larger of the
two classes (Prinzie and Van den Poel, 2008). Failure on road pavements is avoided with the
implementation of proactive maintenance and, therefore, the reality of road pavement data
results in an unbalanced dataset. To offset this imbalance, and subsequently remove any
biasness from the predictions of the trained model, a weighting factor was applied to the data.
Following Equation 8-3, the weighting factor of a datapoint is based on the ratio of non-
− ∑ୀଵ
#
Equation 7-3
( = − 1
× # + 1
∑ୀଵ
#
= !
It is good practice to reserve a portion of the entire dataset for testing (Gil and Johnnson,
2011), if possible, although researchers have assorted viewpoints on the size of the evaluation
197
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
The datasets for the majority of the studies listed in Table 8-1 were sufficient in size to
reserve an adequately sized evaluation dataset to assess the performance of the models. In
some instances, the evaluation dataset was in excess of 30,000 samples (Caruana and
Niculescu-Mizil, 2006; Prinzie and Van den Poel, 2008). The studies with smaller testing sets
insufficient amount of data available for the study. Alternative sampling techniques are
available (Raudys, 2001; Roiger and Geatz, 2003) if the ideal volume of data is not available:
used in training the model. The training of the model and testing is repeated n
198
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
times, once on each test subsample, and the results are averaged over the number
of repetitions;
• Bootstrapping: The training set is sampled with replacement from the complete
testing and the remainder of the dataset is used for training. This method will be
The State Highway LTPP research dataset did not contain sufficient datapoints for separate
training and evaluation datasets, unlike the studies in alternative research domains identified
in Chapter Two and Table 8-1 above. The need for training computational models on small
datasets is a requirement for road pavement data. However, this research dataset was of
adequate size to disregard the leave-one-out cross-validation method. Although Dreiseitl and
method is often overly optimistic despite its simplicity. Therefore, a 10-fold cross-validation
sampling method was adopted in this research to ensure validity of the results of the
performance assessment of the techniques. To achieve this, the entire dataset was randomly
divided into 10 subsamples, with each 10 % subsample reserved once for testing the model
that had been trained on the remaining 90 % of the data. However, extensive run time was
required to repeat the training and testing of the model, resulting in n-times the single
computation time to evaluate the method. The randomness of the sampling method did not
guarantee the same testing dataset was used for each technique; however, the outcome of the
199
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
comparative study was not negatively influenced by this given the method of hypothesis
testing employed.
Point and integrated performance measures were identified in the literature review to evaluate
the performance of the classification techniques, in conjunction with the model accuracy.
Since accuracy is greatly influenced by the class distribution of the dataset, sole reliance on
this performance measure results in a bias assessment of the technique’s predictive power
(Ben-David, 2008; Dreiseitl and Ohno-Machado, 2002; Parker, 2011). Thus, alternative
performance measures were included in the comparative study, specifically point measures
techniques (Parker, 2011), as these calculate the predictive power of the model over all
possible thresholds. To report on such measures, such as the popular area under the receiver
operating characteristic curve (e.g. integration functions), the user requires a cost function
(Dreiseitl and Ohno-Machado, 2002). This research assumed the cost to be uniform between
the predictions that incorrectly predicted failed and sound pavements, such that it is equally
undesirable for false negatives and false positives; therefore, calculating such measures
measure would add no additional value to the results. The consequence of failure varies
depending on the required maintenance treatment. For example, the costs, both financial and
social, of a full rehabilitation are much more extensive than a resealing exercise. Without this
information, the assumption adopted in this research was considered to be adequate, given the
available data. For a more accurate representation of the performance of the classifier,
200
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
Chapter Six presented the development of the prototype system, which, by following the
conceptual design of the research, successfully inferred engineering knowledge through the
use of failure charts (developed in Chapter Four) into the computational model (selected in
The conceptual design of this research focussed on inferring engineering knowledge into the
design of the computational system, such that the development of the support vector
machines models were informed by the developed failure charts. Modelling each successful
factor combination individually, selected by 100 % accuracy (and 90 % in the case of shear
failure), represented a possible failure path on the developed failure charts. Although these
factor combinations correlated well with the failure charts, the use of computational outputs
identifying the failure paths has the potential of excluding critical failure paths, especially if
the factor combinations are not well represented in the data. This hinders the transferability of
the developed prototype system to other road networks; however, the resulting failure paths
included in the system were considered appropriate for the State Highway LTPP dataset.
As each failure path (factor combination) is modelled separately, the computational model
excluded possible interactions between the failure paths, assuming the critical failure path
201
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
was solely responsible for failure. The failure charts account for the interactions between the
causes of failure, as shown per failure path; however, there was no provision included in the
system for interactions between the failure paths of each failure mechanism (refer Section
8.4.1).
Chapter Six demonstrated the difficulties of representing real-life pavement failures with
mathematical equations. The occurrence of two or more failure mechanisms was dealt with
1. Assumed the probability of overall failure did not increase with the occurrence of
2. Assumed independence between the failure mechanisms, such that the overall
failure probability accounted for multiple failures but secondary effects were not
The limitations of each approach were discussed in Section 6.5.2. Although these two
approaches were considered sufficient given the available data, a third approach was
proposed, which was proposed to determine the probabilities of multiple failures using
number of observations where multiple failure modes were apparent and detailed information
regarding the timing of these failure(s), in order to differentiate between secondary effects
and combined (or multiple) failures. Such data and information on these types of failure is
seldom included in road network data, not to say such failures do not occur on the networks.
In most cases, a secondary defect is often masked by the occurrence of primary failure, or so
determined by the site inspection, and subsequently the maintenance fails to address the
202
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
secondary failure causes. This research however used alternative approaches to compute the
Notwithstanding, the expected differences between the two approaches adopted in this
research were apparent in the results in Section 6.6; naturally greater probabilities were
yielded by the probability theory (probability equation) approach, due to the addition of the
comparison of the two approaches failed to conclude a superior approach, based on the
performance measures, although Figures 6-5, 6-6 and 6-7 showed a discernible difference in
the probability distributions. Integrated performance measures would assess this difference
sufficiently if the consequences of incorrect predictions were known and, therefore, provide a
Chapter Seven demonstrated the use of the developed prototype system and evaluated the
effectiveness of the system given an independent testing dataset. The Local Authority LTPP
dataset included identical parameters to the research dataset enabling it to be analysed by the
prototype system. However, typical road network datasets lack the sophistication and detail
included in the LTPP datasets and, therefore, it should not be assumed that the results in
Chapter Seven are equivalent to a network test. Such differences in the datasets were evident
in the preliminary data analysis of the Southland District Council road network and the LTPP
The applications demonstrated in Chapter Seven rely heavily on the predictions from the
system. However, if data needed for the implementation of the prototype system is absent, as
a result of discernible differences between data collection methods and collation of the
203
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
research dataset and RCAs, it is expected that the performance of the prototype system would
further decrease. Consequently, the information obtained from the analysis and manipulation
of the raw probabilities will be ambiguous to the user. To address this, the system could be
retrained on the available data, yet the performance of the new computational models would
need to be evaluated.
The previous sections of this chapter have discussed the issues with the available research
• Restriction of testing methods given the size of the research dataset (refer Section
8.4.2.4), and
Section 8.4.2.5).
Furthermore, this research chose to exclude various data items from the study given the
categorical nature17 of the information and / or the high level of incompleteness. The failure
charts (Figures 4-7 to 4-11) developed in Chapter Four included causes of failure that, due to
the lack of corresponding data, were unable to be included in the classification models. This
gradient, and inadequate compaction. The LTPP datasets listed material types as categorical
data with many possibilities, if the information was available. Given the subjective nature of
17
Data items listed as categories or classes, such as material types, as opposed to numerical data.
204
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
information on such data items poses difficult with current data collection methods and
therefore was not available for this research. As a result, the computational models ignored
any irrelevant, unavailable, or categorical data. This research suggests the development of a
robust methodology for the collection of subjective causes of failure for further consideration
As stated previously (refer Section 8.4.4), the differences between the research dataset and
network data are discernible. Chapter Four identified the failure factors involved in road
pavement failure. Further to this, Chapter Six concluded on the successful factor
combinations from the comparative study presented in Chapter Five. As Table 8-2 shows,
this research concluded strength, composition, and traffic factors were the influential in
predicting the probability of each pavement failure mechanism. These factors should be
collected robustly from road networks, for the success of the prototype system in road
Table 7-6: Summary of the Factor Combinations included in the Prototype System
The performance of the developed prototype system for analysing the failure probabilities of
the State Highway LTPP road network was discussed in Chapter Six. As reported, the
205
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
performance of the shear support vector machines system was surpassed by the other two
failure mechanisms. This was a direct result from excluding the information about material
compositions from the research dataset. Further to the above, such information is required for
Despite the reported success of the support vector machines technique in the comparative
study, the weak results from the phi coefficient suggested little correlation between the input
variables and predicted outputs. The inferred knowledge from the failure paths, although
successful for rutting and fatigue cracking, was not as successful in the shear failure system.
However, given the other reported performance measures, the developed system could be
considered to have performed adequately overall, with respect to predicting rutting and
fatigue cracking.
pavement data;
• Predicting the probability of road pavement failure with support vector machines.
Failure Charts:
Representing knowledge of road pavement failure in the form of failure charts was not
evident in the literature review. By presenting the causes of failure on failure charts, the
causes, including the combinations of causes, are easily identified for future reference and
206
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
added a diagnostic element to the computational model. This research based the development
of the failure charts for rutting, cracking, and shear road failure mechanisms on approaches
previously employed for other infrastructure assets. This work has been submitted to the
Australian Road Research Board Road and Transport Research journal for publication
Comparative Study:
The literature identified one study (Chandra et al., 2009) that had compared the classification
techniques considered for the prototype system in Chapter Five; however, the focus of this
study was outside of the transportation sector. Comparative studies of classification methods
were not evident in the literature within the transportation sector and, more specifically, for
predicting the end of life probability of an infrastructure asset. Methodologies from the
for the future use of classifiers. The performance of the suitable classification techniques was
reported in Chapter Five and this research concluded the appropriateness of classifiers when
The preceding two elements of the design were used as the foundation of the prototype
design, the inference of engineering knowledge in this research, such that the development of
the computational model is informed by the failure charts, is a new approach to modelling
pavement performance and the probability of pavement failure. In doing so, this research
recognised data factors that played an important role in the prediction of pavement failure,
which further identified the importance of strength, composition, and traffic data in predicting
207
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
Using support vector machines to predict the probability of road pavement failure is a novel
approach in the transportation sector. The majority of the literature on pavement performance
end of asset life failure. Schlotjes et al. (2013a) presented the appropriateness of using
support vector machines in the transportation industry and the work was further submitted to
This chapter has critically reviewed the research methodology adopted in the development of
a prototype system for predicting the probability of road pavement failure. In particular, the
effectiveness of the system and assumptions made throughout the research were discussed
and, where appropriate, suggestions have been offered to facilitate future developmental
The inference of engineering knowledge into the computational model was achieved through
the use of failure charts. Although the overall system performed well, the assumptions made
in this research affected the ability of shear failures to be predicted with as great an accuracy
to the other two modes of failure considered. The developed failure charts informed the
development of the computational models, such that there was a parallel association between
the failure charts and model. However, the exclusion of categorical and unattainable data
from the modelling process resulted in a number of the possible failure paths having to be
disregarded.
208
Chapter Eight: SUMMARY AND DISCUSSION
A Review of the Research
Furthermore, suggestions for further development of the system included evaluating the
interactions between the failure mechanisms with the use of computational techniques on
sufficient road pavement data, evaluating the modelling technique with integrated
performance measures, and improvements to the current data collection processes for such
tasks.
Further testing on typical road network datasets is recommended. The study concluded that
strength, composition, and traffic failure factors are essential in obtaining sufficient
predictions with the prototype system. Therefore, for the implementation of the developed
prototype system to other road network datasets, it is imperative that RCAs employ robust
data collection and collation processes in order to attain the input variables required for the
system.
The value of this research, focussing on predicting the end of life probability for road
pavements, was discussed with reference to the literature. As studies in this topic area were
not evident, the use of support vector machines in predicting the probability of pavements is a
Conclusions from the research together with recommendations for future work are presented
209
210
Chapter Nine
CONCLUSIONS AND
RECOMMENDATIONS
As discussed in the previous chapter, this research has demonstrated the objectives outlined
1. Developing failure charts for rutting, cracking, and shear road pavement failure
failure;
211
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
3. Developing a prototype system based on the outcomes from the above two
particularly useful in evaluating the methods for both the performance and the
implementation of the techniques for the given research dataset, despite the
secondary effects of others. This research not only yielded results that confirmed
that support vector machines, from those methods considered, was the most
212
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
proved to be more suitable because of its superior model format and decision
boundary versus the limited expressiveness and simplicity of the model format
for the independent testing dataset, as shown in Chapter Seven. However, using
the prototype system without proper calibration on other road datasets can
compromise the performance of the system. For the dataset tested in this research,
management tool to determine the failure distribution across the network, the
Four add a diagnostic element to the predictions enabling the most probable
presented in Schlotjes et al. (2011). Purely data driven processes are successful in
situations where large amounts of unbiased data is available; however, the results
will undoubtedly be resultant on the trends and relationships within the data. On
213
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
• The success of any computational model depends greatly on the quality of the data
available to train the model with. The inputs included in the prototype system
reflect the data collected in the LTPP programme and, given the detail included in
While the results presented in Chapter Seven have demonstrated the promise of the system, it
is recommended that in order for the developed support vector models to be implemented in
New Zealand, the model needs to be calibrated and tested on a wide variety of different road
networks following the procedure described in Chapter Seven. At the time of submission of
this research, such testing was being facilitated on a variety of New Zealand road networks,
To further develop and improve the accessibility of the prototype system to practitioners, it is
recommended that a computer language, appropriate for the software packages typically
employed in the industry, be developed to call in the support vector machines model files
(currently in a .txt file format in R language) for the direct implementation of the developed
system. For the implementation of the system in a domain similar to that reported in this
the prototype system, the network dataset can be passed through the uploaded model files,
providing the inputs from the network dataset correspond to those of the system. However,
for alternative road networks comprising of pavement compositions and traffic volumes other
214
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
than those considered in this research, it is recommended that new support vector machines
model files are developed, specific for the new dataset, following the methodology adopted
The methodology adopted in this research can be transferred to other road pavement types for
the development of similar pavement performance systems; however, the assumptions of this
research may not be suitable to other pavement domains. Internationally, pavement types and
construction practices differ, such that alternative causes are responsible for failure of other
pavement types than the knowledge inferred in this research, which captured the causes of
flexible road pavement failure. Therefore, it is suggested to reconsider the causes included in
the failure charts for true replication of alternative pavement failure. Although this research
concluded support vector machines were the most appropriate classification technique for this
case study, the performance of this technique may not be reproduced when using alternative
road pavement datasets. Therefore, this research further recommends adopting the
A summary of the recommendations for the improvement of the prototype system are listed
in the following sections, namely addressing the limitations of the research and for further
work.
Chapter Eight discussed the limitations of the research; therefore, the following
215
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
failure may not have been represented in the failure charts. The methodology included site-
specific failure causes; however, an analysis of a greater number of New Zealand road
take into account the interactions between the failure mechanisms in the failure charts.
Comparative study:
To improve the comparative study, this research recommends evaluating the classification
addition to the reported results. To do so, it is important that the research dataset includes
information regarding the cost functions of incorrect predictions for the evaluation of the
It is also recommended a cost effective analysis is undertaken, such as reported in (Levin and
In the design of the prototype system, basing the selection of successful factor combinations
on 100 % accuracy threshold (or 90 % for the case of the shear models) may have excluded
possible failure paths from the resultant system. For example, the factor combination(s)
presenting an accuracy of 99 % were omitted from the remainder of the study, despite the
minute difference between the accuracy and the target threshold. Reviewing this approach, it
is suggested to cross-reference the excluded failure paths with the failure charts to ensure the
216
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
omitted failure paths are not in fact possible for the respective failure mechanism. One
method to address this would be to reduce the accuracy threshold, for example.
Multiple failure events should also be modelled with the support vector machines technique
This research tested the developed prototype system on an independent dataset in Chapter
Seven; however, this testing dataset is atypical of network data. Although the results in
Chapter Seven suggested that the developed rutting system is compromised when used with
the system on typical road networks. It is suggested that the network datasets involved in the
further testing of the system include as many input variables possible. Those associated with
the strength of the pavement and its composition, as well as reliable traffic data, were shown
Possible improvements to the research dataset include information on the costs of failures for
of multiple failures, processes to collate and quantify subjective and categorical data, and
greater amounts of data representing multiple failures to use in modelling such failure types.
217
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
The following recommendations are suggested to improve the developed prototype system:
Chapter Two reviewed the advantages and disadvantages of a list of discriminative and
generative classifiers, yet the comparative study in Chapter Five investigated the performance
of only five methods. The selection of these methods was based on objective criteria,
assessing the applicability and suitability of each of these methods to the study. However, the
exclusion of the other methods (refer Section 2.6) resulted in the unknown performance of
these to the research dataset. Therefore, this research suggests evaluating the excluded
The scope of this research limited the development of the prototype system to three
predominant structural failure mechanisms on New Zealand road networks. Other failure
modes, generally associated with the surfacing layers, impede the function and integrity of
road pavements. For a complete failure system, the framework of this research should be
Employing pavement deterioration models to forecast the future condition state of the
the current PMS practices and failure knowledge base, this research should be further
developed so to include forecasting capabilities in the design of the prototype system; so over
time, a measurable shift in network failure probabilities can be used in future maintenance
decisions.
218
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
Uncertainty:
Comparisons of the research dataset and RCA network datasets identified a number of
inconsistencies. The limited resources and funding available to RCAs restricts the data
collected in road inspections, the collection methods employed, and the timings of road
inspections. Further research is proposed to evaluate the confidence around the predicted
probabilities by incorporating an uncertainty element into the system, taking into account:
This research has proposed a novel approach to predicting the probability of road pavement
failure, contributing to the existing knowledge base of pavement performance models. The
most significant lessons learnt, based on the findings of this research, include:
highlighted in Chapter Five, and revisited in Chapter Six, correlated well with the
developed failure charts from Chapter Four, since the model foundations were
based on the failure paths. The combination of the two elements resulted in a well
219
Chapter Nine: CONCLUSIONS AND RECOMMENDATIONS
• Improvements in the current pavement data are required for effective and
220
REFERENCE LIST
Abraham, D.M., Wirahadikusumah, R., Short, T.J. and Shahbahrami, S. 1998. Optimization
Aitkin, M. and Foxall, R. 2003. Statistical modelling of artificial neural networks using the
multi-layer perceptron. Statistics and Computing, Vol. 13, Issue 3, pp. 227-239.
Anghel, C.I. 2009. Risk assessment for pipelines with active defects based on artificial
intelligence methods. International Journal of Pressure Vessels and Piping, Vol. 86,
221
REFERENCE LIST
Arambula, E., George, R., Xiong, W. and Hall, G. 2011. Development and validation of
Arampamoorthy, H. and Patrick, J.E. 2010. Failure probability of New Zealand pavements.
Agency.
needs using logistic models. Journal of Infrastructure Systems, Vol. 7, No. 4, pp. 160-
165.
Austin, P.C., Tu, J.V. and Lee, D.S. 2010. Logistic regression had superior performance
hospitalized with heart failure. Journal of Clinical Epidemiology, Vol. 63, pp. 1145-
1155.
Austroads. 2006. Guide to Asset Management: Part 5E: Cracking. Sydney, Australia:
Austroads Inc.
Austroads. 2007a. Guide to Asset Management: Part 5B: Roughness. Sydney, Australia:
Austroads Inc.
Austroads. 2007b. Guide to Asset Management: Part 5C: Rutting. Sydney, Australia:
Austroads Inc.
222
REFERENCE LIST
Austroads. 2009b. Guide to Asset Management: Part 5H: Performance Modelling. Sydney,
Ayyub, B.M. and McCuen, R.H. 2003. Probability, Statistics, and Reliability for Engineers
and Scientists. Second Edition. Florida, United States of America: Chapman and
Baesens, B., Verstraeten, G., Van den Poel, D., Egmont-Peterson, M., Van Kenhove, P. and
Vanthienen, J. 2004. Bayesian network classifiers for identifying the slope of the
Baik, H-S., Jeong, H.S. and Abraham, D.M. 2006. Estimating transition probability in
Journal of Water Resources Planning and Management, Vol. 132, Issue 1, pp. 15-24.
Bailey, R., Patrick, J.E. and Jackett, R. 2006. Relationship between design and predicted
Bapat, R.B. 2000. Linear Algebra and Linear Models. Second Edition. New York, United
Basheer, P.A.M., Chidiac, S.E. and Long, A.E. 1996. Predictive models for deterioration of
concrete structures. Construction and Building Materials, Vol. 16, No. 1, pp. 27-37.
223
REFERENCE LIST
Bell, D.D., Cox, L., Jackson, S. and Schaefer, P. 1992. Using casual reasoning for automated
failure modes and effects analysis. Presented at the Annual Reliability and
Ben-Akiva, M. and Gopinath, D. 1995. Modeling infrastructure performance and user costs.
Ben-David, A. 2008. About the relationship between ROC curves and Cohen’s kappa.
using genetic algorithms. Paper No. 980227. Pennsylvania, United States of America:
Bhattacharya, B. and Solomatine, D.P. 2006. Machine learning in soil classification. Neural
Bicego, M., Pȩkalska, E., Tax, D.M.J. and Duin, R.P.W. 2009. Component-based
discriminative classification for hidden Markov models. Pattern Recognition, Vol. 42,
Cai, K-Y. 1996. System failure engineering and fuzzy methodology: An introductory
overview. Fuzzy Sets and Systems, Vol. 83, Issue 2, pp. 113-133.
224
REFERENCE LIST
Chandra, D.K., Ravi, V. and Bose, I. 2009. Failure prediction of dotcom companies using
hybrid intelligent techniques. Expert Systems with Applications, Vol. 36, Issue 3, Part 1,
pp. 4830-4837.
Chen, M., Zheng, A.X., Lloyd, J., Jordan, M.I. and Brewer, E. 2004. Failure diagnosis using
Committee of State Road Authorities (CSRA). 1992. TMH 9: 1992: Pavement Management
System: Standard Visual Assessment Manual for Flexible Pavements. Pretoria, South
Africa.
Contini, S. and Matuzas, V. 2011. Analysis of large fault trees based on functional
decomposition. Reliability Engineering and System Safety, Vol. 96, Issue 3, pp. 383-
390.
Cortes, C. and Vapnik, V. 1995. Support-vector networks. Machine Learning, Vol. 20, Issue
3, pp. 273-297.
Creagh, M. 2005. Risk assessment for unbound granular materials and pavement
Brisbane, Australia.
Davies, S. and Ihaka, R. 2010. Fitting generalized linear models. R Documentation. [Online].
Deng, Z. and Henning, T.F.P. 2012. Testing of new NZTA condition performance measurers.
225
REFERENCE LIST
Dimitriadou, E., Hornik, K., Leisch, F., Meyer, D. and Weingessel, A. 2011. Package
Dreiseitl, S. and Ohno-Machado, L. 2002. Logistic regression and artificial neural network
Eastaugh, J.L., Smye, S.W., Snowden, S., Walker, J.J., Dear, P.R.F. and Farrin, A. 1997.
Comparison of neural networks and statistical models to predict gestational age at birth.
Eftekhar, B., Mohammad, K., Ardebili, H.E., Ghodsi, M. and Ketabchi, E. 2005. Comparison
of artificial neural network and logistic regression models for prediction of mortality in
head trauma based on initial clinical data. BMC Medical Informatics and Decision
Emery, S. J. 1992. The Prediction of Moisture Content in Untreated Pavement Layers and an
Roads and Transport Technology (DRTT), Council for Scientific and Industrial
Research (CSIR).
Everitt, B.S. and Hothorn, T. 2006. A Handbook of Statistical Analyses Using R. Florida,
Faraway, J.J. 2006. Extending the Linear Model with R. Florida, United States of America:
226
REFERENCE LIST
Fawcett, T. 2006. An introduction to ROC analysis. Pattern Recognition Letters, Vol. 27,
Foote, J.T. 1994. Decision-tree probability modeling for HMM speech recognition. Rhode
Gao, L., Aguiar-Moya, J.P. and Zhang, Z. 2012. Bayesian analysis of heterogeneity in
Gil, D. and Johnsson, M. 2011. Support Vector Machines in Medical Classification Tasks, in
Boyle, B.H. (Ed.). Support Vector Machines: Data Analysis, Machine Learning and
Applications. New York, United States of America: Nova Science Publishers, Inc.
Gribble, M. and Patrick, J. 2008. Adaptation of the AUSTROADS pavement design guide for
New Zealand conditions. Land Transport New Zealand Research Report 305.
Gualtieri, J.A., Chettri, S.R., Cromp, R.F. and Johnson, L.F. 1999. Support vector machine
classifiers as applied to AVIRIS data. Presented at the 8th Annual JPL Airborne Earth
Gupta, A. and Lam, S.M. 1998. Weight decay backpropagation for noisy data. Neural
Gupta, A., Kumar, P. and Rastogi, R. 1997. Pavement deterioration and maintenance model
for low volume roads. International Journal of Pavement Research and Technology,
227
REFERENCE LIST
Haas, R., Hudson, W.R. and Zaniewski, J.P. 1994. Modern Pavement Management. Malabar,
Halilčević, S.S., Gubina, F. and Gubina, A.F. 2011. The composite fuzzy reliability index of
power systems. Engineering Applications of Artificial Intelligence, Vol. 24, Issue 6, pp.
1026-1034.
Hawkins, P.G. and Woollons, D.J. 1998. Failure modes and effects analysis of complex
Hayward, B.J. 2006. Investigation of road base shear strains using in-situ instrumentation.
Thesis].
He, J., Zhang, Y., Li, X. and Shi, P. 2012. Learning naïve Bayes classifiers from positive and
Henning, T.F.P. 2008. The development of pavement deterioration models on the state
Henning, T.F.P. and Roux, D.C. 2008. Pavement deterioration models for asphalt-surfaced
228
REFERENCE LIST
Henning, T.F.P., Roux, D.C. and Alabaster, D. 2007. Benchmarking pavement performance
between Transit’s LTPP and CAPTIF programs. Land Transport New Zealand
Research Report 319. Wellington, New Zealand: Land Transport New Zealand.
Henning, T.F.P., Costello, S.B., Dunn, R.C.M., Parkman, C.C. and Hart, G. 2004. The
highway network. Road and Transport Research, Vol. 13, Issue 2, pp. 17-32.
Henning, T.F.P., Costello, S.B., Parkman, C.C., Bennett, C.R. and Dunn, R.C.M. 2006. A
complete review of the crack initiation models for New Zealand. Presented at the 22nd
Horak, E. 2008. Benchmarking the structural condition of flexible pavements with deflection
bowl parameters. Journal of the South African Institution of Civil Engineering, Vol. 50,
Hu, H., Du, X., Xu, C., Zhao, F., Lin, X. and Bo, Z. 2011. Risk assessment of cascading
failure in power systems based on uncertainty theory. Presented at the IEEE Power and
Hu, Y-C. 2010. Analytic network process for pattern classification problems using genetic
Hussain, J., Henning, T., Wilson, D.J. and Alabaster, D. 2011. What happens when it rains?
229
REFERENCE LIST
Isa, A.H.M., Ma’soem, D.M. and Hwa, L.T. 2005. Pavement performance model for federal
roads. Presented at Eastern Asia Society for Transportation Studies (EASTS). Bangkok,
Jagielska, I., Matthews, C. and Whitfort, T. 1999. An investigation into the application of
neural networks, fuzzy logic, genetic algorithms, and rough sets to automated
knowledge acquisition for classification problems. Neurocomputing, Vol. 24, Issue 1-3,
pp. 34-54.
Karatzoglou, A., Meyer, D. and Hornik, K. 2006. Support vector machines in R. Journal of
Kaseko, M.S. and Ritchie, S.G. 1993. A neural network-based methodology for pavement
Berlin Heidelberg.
Kleiner, Y., Adams, B.J. and Rogers, J.S. 2001. Water distribution network renewal planning.
Kuncheva, L.I. 2006. On the optimality of naïve Bayes with dependent binary features.
230
REFERENCE LIST
Land Transport New Zealand. 2008. Land Transport New Zealand road network statistics for
http://www.landtransport.govt.nz/performance/2008/docs/network-statistics.pdf.
Lazzaroni, M., Cristaldi, L., Peretto, L., Rinaldi, P. and Catelani, M. 2011. Reliability
Levin, H.M. and McEwan, P.J. 2001. Cost-Effectiveness Analysis: Methods and Applications.
Available: http://cran.r-project.org/web/packages/randomForest/randomForest.pdf.
[December 2, 2010].
Lim, T-S., Loh, W-Y. and Shih, Y-S. 2000. A comparison of prediction accuracy,
complexity, and training time for thirty-three old and new classification algorithms.
Lindhe, A., Rosén, L., Norberg, T. and Bergstedt, O. 2009. Fault tree analysis for integrated
and probabilistic risk analysis of drinking water systems. Water Research, Vol. 43,
Long, P.M., Servedio, R.A. and Simon, H.U. 2007. Discriminative learning can succeed
where generative learning fails. Information Processing Letters, Vol. 103, Issue 4, pp.
131-135.
231
REFERENCE LIST
Long, W., Sato, Y. and Horigome, M. 2000. Quantification of sequential failure logic for
fault tree analysis. Reliability Engineering and System Safety, Vol. 67, Issue 3, pp. 269-
274.
Lounis, Z. and Madanat, S.M. 2002. Integrating mechanistic and statistical deterioration
models for effective bridge management. Presented at the 7th International Conference
States of America.
Lytton, R.L. 1987. Concepts of pavement performance and modelling. Presented at the 2nd
Madanat, S.M., Karlaftis, M.G. and McCarthy, P.S. 1997. Probabilistic infrastructure
deterioration models with panel data. Journal of Infrastructure Systems, Vol. 3, No.1,
pp. 4-9.
Berlin Heidelberg.
Manik, A., Gopalakrishnan, K. and Khaitan, S.K. 2007. Statistical modeling of accelerated
Computer, Information, and Systems Science, and Engineering, Vol. 1, Issue 1, pp. 59-
67.
Martin, T.C. 2008. Predicting sealed granular pavement deterioration at a road network level.
Menard, S. 2002. Applied Logistic Regression Analysis. Second Edition. Sage University
232
REFERENCE LIST
Mendenhall, W. and Beaver, R.J. 1991. Introduction to Probability and Statistics. Eighth
Morcous, G., Rivard, H. and Hanna, A.M. 2002. Case-based reasoning system for modeling
pp. 104-114.
Schoekkopf, B., Tsuda, K. And Vert, J.P. (Eds). Kernel Methods in Computational
Biology. Cambridge, Massachusetts, United States of America: MIT Press, pp. 71-92.
Örkcü, H.H. and Bal, H. 2011. Comparing performances of backpropagation and genetic
algorithms in the data classification. Expert Systems with Applications, Vol. 38, Issue 4,
pp. 3703-3709.
Ortmeier, F. and Schellhorn, G. 2007. Formal fault tree analysis – practical experiences.
Electronic Notes in Theoretical Computer Science, Vol. 185, Special Issue, pp. 139-
151.
Osman, H. and Bainbridge, K. 2011. Comparison of statistical deterioration models for water
3, pp. 259-266.
233
REFERENCE LIST
International Journal for Numerical and Analytical Methods in Geomechanics, Vol. 30,
Pan, N-F., Ko, C-H., Yang, M-D. and Hsu, K-C. 2011. Pavement performance prediction
through fuzzy regression. Expert Systems with Applications, Vol. 38, Issue 8, pp.
10010-10017.
Pan, N-F., Lin, T-C. and Pan, N-H. 2009. Estimating bridge performance based on a matrix-
driven fuzzy linear regression model. Automation in Construction, Vol. 18, Issue 5, pp.
578-586.
Papagiannakis, A.T. and Masad, E.A. 2008. Pavement Design and Materials. New Jersey,
Park, D-Y., Buch, N. and Suh, Y-C. 2001. Development of fatigue cracking prediction model
for flexible pavements. KSCE Journal of Civil Engineering, Vol. 5, No. 4, pp. 397-402.
Parker, C. 2011. An analysis of performance measures for binary classifiers. Presented at the
Patev, R.C., Putcha, C. and Foltz, S.D. 2005. Methodology for Risk Analysis of Dam Gates
and Associated Operating Equipment Using Fault Tree Analysis. United States of
234
REFERENCE LIST
Peng, C-Y.J., Lee, K.L. and Ingersoll, G.M. 2002. An introduction to logistic regression
analysis and reporting. The Journal of Educational Research, Vol. 96, No. 1, pp. 3-14.
Perlich, C., Provost, F. and Simonoff, J.S. 2003. Tree induction vs logistic regression: A
211-255.
Pickard, K., Müller, P. and Bertsche, B. 2005. Multiple failure mode and effects analysis –
Powers, D.M.W. 2011. Evaluation: From precision, recall and F-measure to ROC,
Prinzie, A. and Van den Poel, D. 2008. Random forests for multiclass classification: Random
multinomial logit. Expert Systems with Applications, Vol. 34, Issue 3, pp. 1721-1732.
Prozzi, J.A. and Madanat, S.M. 2002. A nonlinear model for predicting pavement
R Core Team. 2011. Wilcoxon rank sum and signed rank tests. R Documentation. [Online].
Available: http://stat.ethz.ch/R-manual/R-patched/library/stats/html/wilcox.test.html.
[February 2, 2011].
235
REFERENCE LIST
Rau, C.A., Becker, D.G., Besuner, P.M., Cipolla, R.C., Egan, G.R., Gupta, P., Johnson, D.P.,
Omry, U., Tetelman, A.S., Rettig, T.W. and Peters, D.C. 1977. Failure analysis and
failure prevention in electric power systems. Nuclear Engineering and Design, Vol. 43,
Razi, M.A. and Athappilly, K. 2005. A comparative predictive analysis of neural networks
(NNs), nonlinear regression and classification and regression tree (CART) models.
Reay, K.A. and Andrews, J.D. 2002. A fault tree analysis strategy using binary decision
diagrams. Reliability Engineering and System Safety, Vol. 78, Issue 1, pp. 45-56.
Reed, R.D. and Marks II, R.J. 1999. Neural Smithing. Massachusetts, United States of
using risk analysis. West Virginia, United States of America: West Virginia University
[Ph.D. Thesis].
for conversion of fault trees to binary decision diagrams. Reliability Engineering and
236
REFERENCE LIST
Birmingham.
Rogers, S. and Girolami, M. 2012. A First Course in Machine Learning. Florida, United
Roiger, R.J. and Geatz, M.W. 2003. Data Mining: A Tutorial-Based Primer. United States of
Sadek, A.W., Freeman, T.E. and Demetsky, M.J. 1996. Deterioration prediction modeling of
Virginia’s interstate highway system. Transportation Research Record, Vol. 1524, pp.
118-129.
Saghafi, B., Hassani, A., Noori, R. and Bustos, M.G. 2009. Artificial neural networks and
Salt, G., Henning, T.F.P., Stevens, D. and Roux, D.C. 2010. Rationalisation of the structural
Transport Agency.
237
REFERENCE LIST
Schlotjes, M.R. and Henning, T.F.P. 2012. Knowing your risk envelope – An enhanced
understanding of the health of road networks. Presented at the 25th Ingenium Annual
Schlotjes, M.R., Henning, T.F.P. and Burrow, M.P.N. 2011. Improved network
Schlotjes, M.R., Henning, T.F.P. and Burrow, M.P.N. 2012a. The development of a
diagnostic approach that predicts failure risk of flexible road pavements. Presented at
Schlotjes, M.R., Henning, T.F.P. and St George, J.D. 2009. The risks associated with
pavement failure in New Zealand. Presented at NZIHT and NZTA 10th Annual
Schlotjes, M.R., Burrow, M.P.N., Evdorides, H.T. and Henning, T.F.P. 2013a. A
methodology to assess the causes and probability of road pavement structural failure.
Schlotjes, M.R., Burrow, M.P.N., Evdorides, H.T. and Henning, T.F.P. 2013b. Using SVM to
238
REFERENCE LIST
Schlotjes, M.R., Henning, T.F.P., Burrow, M.P.N. and St. George, J.D. 2012b. Incorporating
for publication.
system failure mode and effects analysis by decision making trial and evaluation
laboratory technique. Reliability Engineering and System Safety, Vol. 91, Issue 8, pp.
872-881.
Sharabah, A., Setunge, S. and Zeephongsekul, P. 2006. Use of Markov chain for deterioration
Engineering. Singapore.
classification tasks. Information Processing and Management, Vol. 45, Issue 4, pp.
427-437.
Son, C-S., Kim, Y-N., Kim, H-S., Park, H-S. and Kim, M-S. 2012. Decision-making model
for early diagnosis of congestive heart failure using rough set and decision tree
Community.
239
REFERENCE LIST
Telford Limited.
Tso, S.K., Gu, X.P., Zeng, Q.Y. and Lo, K.L. 1998. An ANN-based multilevel classification
approach using decomposed input space for transient stability assessment. Electric
Tu, J. 1996. Advantages and disadvantages of using artificial neural networks versus logistic
regression for predicting medical outcomes. Journal of Clinical Epidemiology, Vol. 49,
Van Looy, S., Verplancke, T., Benoit, D., Hoste, E., Van Maele, G., De Turck, F. and
concentration in liver transplantation patients in the intensive care unit through support
Verikas, A., Gelzinis, A. and Bacauskiene, M. 2011. Mining data with random forests: A
survey and new results of new tests. Pattern Recognition, Vol. 44, Issue 2, pp. 330-349.
Volkanovski, A., Čepin, M. and Mavko, B. 2009. Application of the fault tree analysis for
assessment of power system reliability. Reliability Engineering and System Safety, Vol.
Wang, J., Ruxton, T. and Labrie, C.R. 1995. Design for safety of engineering systems with
multiple failure state variables. Reliability Engineering and System Safety, Vol. 50,
240
REFERENCE LIST
Wang, K.C.P., Zaniewski, J. and Way, G. 1994. Probabilistic behavior of pavements. Journal
Wirahadikusumah, R., Abraham, D.M. and Iseley, T. 2001. Challenging issues in modeling
77-84.
Wu, Y., Ianakiecv, K. and Govindaraju, V. 2002. Improved k-nearest neighbor classification.
Xiao, N., Huang, H.Z., Li, Y., He, L. and Jin, T. 2011. Multiple failure modes analysis and
weighted risk priority number evaluation in FMEA. Engineering Failure Analysis, Vol.
Xu, K., Tang, L.C., Xie, M., Ho, S.L. and Zhu, M.L. 2002. Fuzzy assessment of FMEA for
engine systems. Reliability Engineering and System Safety, Vol. 75, Issue 1, pp. 17-29.
Yang, J., Gunaratne, M., Lu, J.J. and Dietrich, B. 2005. Use of recurrent Markov chains for
Yang, J., Lu, J.J., Gunaratne, M. and Dietrich, B. 2006. Modeling crack deterioration of
Younis, R. and Knight, M.A. 2010. A probability model for investigating the trend of
241
REFERENCE LIST
Yuhua, D. and Datao, Y. 2005. Estimation of failure probability of oil and gas transmission
pipelines by fuzzy fault tree analysis. Journal of Loss Prevention in the Process
Zhong, D., Sun, Y. and Li, M. 2011. Dam break threshold value and risk probability
assessment for an earth dam. Natural Hazards, Vol. 59, Issue 1, pp. 129-147.
242
APPENDICES
A research site visit to the Southland District, New Zealand, was completed between the
dates of 20th April 2009 and 23rd April 2009. The purpose of this site visit was to:
• Gain experience with characteristic RCA road data specifically data from the
RAMM database.
The information from this site visit aided the development of the failure charts. Appendix A
243
APPENDICES
Appendix B.1 provides the results from the preliminary data analysis using the Southland
District Council road dataset. The data included in this dataset was obtained primarily from
the RAMM database, and provides this research with data characteristically typical of RCA
databases. This analysis aided in the development of the specific failure charts for rutting,
Data from the LTPP programme established two datasets for this research. Appendix B.2
presents the results from the preliminary data analysis using both the Local Authority (LA)
and State Highway (SH) LTPP datasets. This analysis aided in the development of the failure
charts in Chapter Four and furthermore the State Highway LTPP analysis was used as a
244
APPENDICES
The computational code employed in the R statistical programme throughout this research is
presented below. The source files for each modelling technique are included in the Appendix
Further to the above, the following (Appendices C.1 to C.5) detail the individual code for
The glm() function (Davies and Ihaka, 2010) was used to construct a logistic regression
model in R, and a number of the arguments used in this function remained as the default
arguments. The following shows the non-default arguments defined in this research:
where:
formula = variable description of the model to be fit
data = dataset used to train the model, which contains all the variables included in the
‘formula’ argument
weights = weighting vector to account for the imbalance of the failures versus non-failures of
the research dataset
245
APPENDICES
The family argument described the error distribution and link function used in the model
where, within this function, the binomial family was chosen as this permits the function for
The predict() function below was used to predict the probability of failure; the type argument
where:
model= the trained logistic regression model, from above
test_data= the testing dataset to assess the modelling technique, and contains all the
variables included in the model training
The nnet package in R (Ripley, 2012) was used to construct a feed-forward neural network
data=data, weights=wgts)
where:
formula = variable description of the model to be fit
data = dataset used to train the model, which contains all the variables included in the
‘formula’ argument
weights = weighting vector to account for the imbalance of the failures versus non-failures of
the research dataset
The size argument defined the number of hidden units (nodes) in the hidden layer and, in this
research, a range of 1 to 10 nodes inclusive were trialled. An arbitrary value for decay was
selected to lower the magnitude of the weights, resulting in a more optimal solution. The skip
argument was set to TRUE to allow the network to skip a layer, if necessary, so the optimal
solution can connect straight from the input nodes to the output layer. Hess was also set to
246
APPENDICES
TRUE so the measure of fit at the best set of weights found was returned in the output
summary. The maximum number of iterations for the training of the model was set to 1000,
using maxit, and entropy was set to TRUE to fit maximum conditional likelihood and provide
probabilistic outputs.
The predict() function was used in the same manner as the logistic regression technique
previously, with the type argument omitted, as the probability output was already defined in
predict(model, test_data)
where:
model= the trained neural network model, from above
test_data= the testing dataset to assess the modelling technique, and contains all the
variables included in the model training
The svm() function within the e1071 package (Dimitriadou et al., 2011) in R was used to
construct the support vector machines model. The research chose the radial basis function
kernel, defined by the kernel argument, because of its ability to handle non-linear attributes
and class relationships (Gil and Johnsson, 2011). The type argument was set to “C” to
indicate a classification type of model, and the inclusion of the probability argument applies
the maximum conditional likelihood concept, alike neural networks, to the decision values.
where:
formula = variable description of the model to be fit
data = dataset used to train the model, which contains all the variables included in the
‘formula’ argument
weights = weighting vector to account for the imbalance of the failures versus non-failures of
the research dataset
247
APPENDICES
Like the previous methods, the predict() function was used to establish the predictions of the
where:
model= the trained support vector machines model, from above
test_data= the testing dataset to assess the modelling technique, and contains all the
variables included in the model training
In R, the rpart package and rpart() function (Themeau and Atkinson, 2012) was used to
construct a probability tree model . The method argument of this function defined the
where:
formula = variable description of the model to be fit
data = dataset used to train the model, which contains all the variables included in the
‘formula’ argument
weights = weighting vector to account for the imbalance of the failures versus non-failures of
the research dataset
Pruning unnecessary branches optimised the model output; therefore, in the same package,
the prune() function was applied to the model tree. This research selected the complexity
parameter (cp) (i.e. the complexity of the tree) that was associated with the smallest cross-
The predict() function was used to establish the predictions for the testing dataset and the
248
APPENDICES
where:
model= the trained probability tree model, from above
test_data= the testing dataset to assess the modelling technique, and contains all the
variables included in the model training
The randomForest() function, from the package of the same name (Liaw and Wiener, 2012),
importance=TRUE, keep.forest=TRUE)
where:
formula = variable description of the model to be fit
data = dataset used to train the model, which contains all the variables included in the
‘formula’ argument
weights = weighting vector to account for the imbalance of the failures versus non-failures of
the research dataset
The xtest was not needed in this research, thus this argument was set as NULL. Throughout
the construction of the forest, the importance of the predictors was assessed, defined by
‘importance = TRUE’. The argument keep.forest=TRUE ensured the forest information was
retained in the output object for future reference. The remainder of the model arguments
Identical to probability trees, the predict() function used the type argument to define the
output as a probability.
where:
model= the trained random forest model, from above
249
APPENDICES
test_data= the testing dataset to assess the modelling technique, and contains all the variables
250
APPENDICES
The results tables are published in Appendix D, which includes individual tables for each of
the cross-validation tests (10 in total for each factor combination), which are then averaged to
provide a single mean value and associated standard deviation for each performance measure
per factor combination, per failure mechanism. In addition to the individual results tables, a
summary table of the averaged cross-validation tests (for each factor combination per failure
mechanism) is presented in this appendix. It is the data from these summary tables that are
used in the density plots and box and whisker plots, also included in this appendix.
The following discusses the difference between the three datasets used in this analysis. This
thesis reports on the results from the NAs=0 State Highway LTPP dataset (as detailed
below):
• Full: The full State Highway LTPP dataset includes all the datapoints from the
LTPP database. As part of the modelling process, missing values (such as NAs)
are removed from the computational models, reducing the number of datapoints
• Limited: This dataset includes only the sites which have failed, and the data from
• NAs=0: A full dataset from the State Highway LTPP database where the missing
values were assigned a default value of zero (0) (as per Chapter Five) to ensure a
251
APPENDICES
A table of successful factor combinations for each failure mechanism (rutting, fatigue
cracking, and shear) is presented in this appendix. These combinations were based on the
success of each, given the success threshold of 100 % accuracy for rutting and fatigue
cracking failure, and 90 % for shear failure. Included in the table is a detailed description of
the factor groups (from Table 4-1) for each combination, and the causes associated with such
failure paths.
252
APPENDICES
• The raw predictions from the State Highway LTPP dataset (please note that
additional data was collected between the years of 2010 – 2011 and added to the
research dataset);
(rutting, fatigue cracking, and shear), and box and whisker plots for each, and
• The raw data tables and plots for the sensitivity analysis, for each failure
mechanism to test the robustness of the prototype system. These results show how
253
APPENDICES
Appendix G contains the result tables from the case study described in Chapter Seven,
including:
• The raw predictions from the prototype system evaluating the probability of
• Tables recording the performance measures for each failure mechanism (rutting,
• The calculation table for the overall failure probabilities for the Local Authority
LTPP rural dataset, based on the two calculation methods discussed in Chapter
Seven.
254